<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by CellCS on Medium]]></title>
        <description><![CDATA[Stories by CellCS on Medium]]></description>
        <link>https://medium.com/@shmilysyg?source=rss-9bd333c8dcb5------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*fID0YS7DuILP4gLgA1fDkg.png</url>
            <title>Stories by CellCS on Medium</title>
            <link>https://medium.com/@shmilysyg?source=rss-9bd333c8dcb5------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 01:15:26 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@shmilysyg/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[From Writing Code to Managing Agents: How Anthropic’s Workforce Is Changing]]></title>
            <link>https://medium.com/@shmilysyg/from-writing-code-to-managing-agents-how-anthropics-workforce-is-changing-a1428161a37c?source=rss-9bd333c8dcb5------2</link>
            <guid isPermaLink="false">https://medium.com/p/a1428161a37c</guid>
            <category><![CDATA[ai-engineer]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[CellCS]]></dc:creator>
            <pubDate>Thu, 12 Nov 2026 01:27:21 GMT</pubDate>
            <atom:updated>2026-11-12T01:27:21.251Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*QFZy-vYH6NcR6mgMwG0TDQ.png" /></figure><p>Anthropic’s Economics team published ‘How employment is changing at Anthropic’ on October 16, 2026 that uses Anthropic itself as a case study of what work looks like when a company goes all-in on AI agents.</p><p><strong>Summary</strong></p><p><strong>Junior technical hiring has dropped sharply.</strong> The share of junior-level employees in technical departments slipped from 21% to 14% since early 2025, and the proportion of junior technical hires fell even faster, from 27% to 9%. By April, engineering and research teams had deemphasized hiring for lower-band technical positions, and around a third of ML managers and a fifth of SWE managers reported hiring less due to AI.</p><p><strong>Agents now do most of the hours.</strong> By August, interactive AI agent hours outnumbered employees’ six to one. Code commits per engineer have risen fastest, at 5x their velocity of six months earlier, and output also rose for lawyers, accounts payable, and sales. When they checked for low-quality “AI slop,” they found no difference in rates of lines of code subsequently removed or reverted between agent-written and human-written code by the same person.</p><p><strong>Human attention is now the bottleneck.</strong> Claude sessions spend around 80% of their time idly waiting for human input. In response, people do “meta-work”: this leapt from 10% of the average employee’s interactive Claude sessions in fall 2025 to 22% in spring 2026, much of it agent management such as writing specs for multi-agent runs, having one agent review another’s output, and monitoring pipelines.</p><p><strong>Coordination got more important, not less.</strong> Even among technical individual contributors, meeting time is up 16% over the past year. Agents need context, so undocumented decisions become a problem: private work in unlogged meetings and direct messages is invisible, confusing and discouraged in an AI-heavy workplace.</p><p><strong>Hiring criteria shifted.</strong> By 2026, mentions of candidates’ creative AI tool use had tripled for technical roles and quadrupled for non-technical roles. Among technical candidates, the biggest increases were in comments on autonomy and judgment, while discussion of domain expertise shrank. One manager’s rationale: “I expect models to continue to be jagged, so for my team I want them to be fungible across domains.”</p><p><strong>Insights</strong></p><p>The most important point is that the job didn’t vanish. It moved up a level of abstraction. Writing code, finding bugs, and drafting emails got automated, and what remained was deciding what to build, specifying it, verifying it, and coordinating with people. One quote captures this well: “that 20% that was strategic becomes 90% of the job.”.</p><p>The junior-hiring drop is the uncomfortable part, and the authors admit their own fixes don’t solve it: employment protections for incumbents will not benefit those not hired to begin with. There’s also a subtle long-term risk they flag: in some roles, people are increasingly removed from the object-level work that originally built their skills. If juniors never do the grunt work, how do they develop the judgment that’s now most valued?</p><p>Engineers are also becoming more like managers. The ratio of engineers to product managers sits around 17 to 1 and has increased since 2025, and commits from managers have doubled. The line between PM, engineer, and manager is blurring.</p><p>Some caveats: Anthropic is an unusual company (selling the AI it uses, growing explosively, with stock-heavy compensation), and adoption wasn’t random, so some productivity gains reflect people who were already ramping up. Treat it as a leading indicator, not a forecast for every workplace.</p><p><strong>Suggestions</strong> for a software developer and AI engineer</p><p><strong>Become good at orchestrating agents, not just prompting them.</strong> The valued skill is running multi-agent workflows: writing clear specs, setting up one agent to review another, building monitoring, and maintaining standing instructions (skill files, CLAUDE.md-style context). Practice this on real projects now, since it’s exactly what hiring managers are starting to look for.</p><p><strong>Build a visible track record of creative AI use.</strong> Interviewers increasingly note this explicitly. A portfolio showing how you built an agent pipeline, an eval harness, or an internal tool that automated a workflow says more than a list of frameworks.</p><p><strong>Invest in judgment and verification skills.</strong> When agents generate 5x the code, the scarce skill is knowing what’s correct, safe, and worth shipping. Get strong at code review, system design, testing strategy, debugging distributed systems, and reading unfamiliar code quickly.</p><p><strong>Stay broad rather than narrowly specialized.</strong> Because models are “jagged,” teams want people who can move across domains. As an AI engineer, being able to span backend, infra, evals, and product thinking is more durable than deep expertise in a single framework.</p><p><strong>Protect your fundamentals deliberately.</strong> Given the skill-atrophy risk, occasionally do hard things by hand: debug without an agent, implement an algorithm yourself, read the source of a library you depend on. That understanding is what lets you catch subtle agent mistakes.</p><p><strong>Work in the open and write things down.</strong> Documenting decisions, design docs, and clear tickets makes you more effective with agents and more visible to teammates. This habit is becoming a core competency rather than a nice-to-have.</p><p><strong>Lean into the human parts.</strong> Stakeholder communication, cross-team alignment, understanding customer needs, and making tradeoff calls are where meeting time is growing. Engineers who can translate between business problems and agent-executable specs will be in demand.</p><p><strong>If you’re early in your career,</strong> the entry path is narrower, so compensate by showing autonomy: ship complete projects end to end (ideally with agents), contribute to open source, and demonstrate that you can own outcomes rather than just tasks.</p><p><strong>Discussion</strong>: meta-work</p><p><strong>Meta-work</strong> is work about the work. Instead of doing a task yourself (writing the function, drafting the contract review), you build and run the system that gets agents to do the task well. The report defines it by what it measured: building Claude skills, plans and reviews for other agents, dashboards and monitors for agent pipelines, and documenting how work is done, in contrast to the actual work itself. It grew from 10% to 22% of employees’ Claude sessions in about six months.</p><p>A simple analogy: a chef cooks dishes; a head chef designs the menu, writes the recipes, trains the line cooks, and tastes before plates go out. Meta-work is the head chef’s job, with agents as the line cooks</p><p>For a software developer, meta-work looks like this:</p><ul><li>Writing a clear spec or plan that several agents can execute in parallel, rather than writing the code yourself.</li><li>Setting up a reviewer agent that checks another agent’s pull requests before you look at them.</li><li>Maintaining standing instructions such as CLAUDE.md files or skills, so agents know your conventions. At Anthropic, these grew from a few hundred in July 2025 to 8,500 a year later.</li><li>Building dashboards that show what your agents are doing and flag what needs human attention.</li><li>Triaging agent requests. One engineer set up an agent “chief of staff” to triage agent requests; another wrote a performance review for their Claude team.</li></ul><p>The reason it exists is the attention bottleneck. Agents produce more output than a person can read, so the only way to scale is to invest in the systems that filter, verify, and direct that output.</p><p><strong>Happy Reading, Happy Thinking, Happy Building.</strong></p><p><strong>References:</strong></p><p><a href="https://www.anthropic.com/research/employment-at-anthropic">How employment is changing at Anthropic web</a></p><p><a href="https://www-cdn.anthropic.com/72628f1863c53432195db83279252dfd2cb53158/Employment%20at%20Anthropic.pdf">How employment is changing at Anthropic pdf</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a1428161a37c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What Comes After Vibe Coding? The Rise of the AI Business Engineer]]></title>
            <link>https://medium.com/@shmilysyg/what-comes-after-vibe-coding-the-rise-of-the-ai-business-engineer-86512c792506?source=rss-9bd333c8dcb5------2</link>
            <guid isPermaLink="false">https://medium.com/p/86512c792506</guid>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[business-strategy]]></category>
            <dc:creator><![CDATA[CellCS]]></dc:creator>
            <pubDate>Thu, 24 Sep 2026 02:05:51 GMT</pubDate>
            <atom:updated>2026-09-24T02:05:51.549Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ZIi2T6B-Pbf0bhIksjYeOQ.png" /></figure><p>A year ago, “<strong>vibe coding</strong>” felt almost magical. Describe an app in plain text with your language. Then let an AI generate the code. Make a few adjustments, deploy, debug, release and REPEAT.</p><p>Today, the landscape is evolving again.</p><p>GPT-6, Claude Opus 5.5, Gemini, DeepSeek, Qwen, and many multimodal models can understand text, code, images, audio, and even real-world workflows. At the same time, coding is no longer just about prompting a chatbot. We now have <strong>agent platforms and harnesses</strong> such as Codex, Meta Muse, DeepSeek Harness, OpenCode, Claude Code, and many others that can plan, code, test, debug, and deploy software with increasing autonomy.</p><p>The question is no longer (The answer is obviously yes):</p><blockquote><em>“Can AI write code?”</em></blockquote><p><strong>The new question may become</strong>:</p><blockquote><em>“</em>Who benefits when everyone can write software<em>?”</em></blockquote><blockquote><em>“</em>What new value is created when everyone can write software?<em>”</em></blockquote><p><strong>1 Vibe Coding Solved the Supply Problem</strong></p><p>For decades, software development had a supply constraint: building software required engineers, Engineers were expensive, Engineering teams were difficult to scale, …..</p><p>AI changed that equation.</p><p>Today, a single developer can build in days what previously required a small team and several months. AI agents can generate code, run tests, fix bugs, and even coordinate with other agents. The industry is already shifting from simple AI-assisted coding toward agentic engineering and multi-agent workflows.</p><p>Software supply is exploding. Apps are everywhere. MVPs appear daily. Landing pages, SaaS products, internal tools, mobile apps, browser extensions, AI agents, and automation workflows can be created faster than ever before.</p><p>But when supply becomes abundant, something else may become scarce:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/738/1*lFoq_nzympfnxiqybnCiUg.jpeg" /><figcaption>Ref:<a href="https://www.linkedin.com/pulse/attention-all-you-need-paper-changed-ai-summarizes-how-sahil-sagar-kjvze">https://www.linkedin.com/pulse/attention-all-you-need-paper-changed-ai-summarizes-how-sahil-sagar-kjvze</a></figcaption></figure><blockquote><strong>Attention!</strong></blockquote><p><strong>2 The Hard Part Is No Longer Building</strong></p><p>Many people still think technology is the bottleneck. Increasingly, it is not.</p><p>The bottleneck is:</p><ul><li>Finding a real problem</li><li>Understanding users</li><li>Distribution</li><li>Marketing</li><li>Trust</li><li>Monetization</li><li>Business operations</li><li>…</li></ul><p>Thousands of AI-generated apps launch every week. Most of them will never get a paying customer. Not because the code is bad. Because nobody needs them, Or nobody knows they exist. Recent discussions around AI entrepreneurship point to a similar conclusion: AI dramatically lowers the cost of building products, <strong>but successful businesses still require insight, product taste, and distribution</strong>.</p><p><strong>Technology is becoming a commodity. Business understanding is becoming the differentiator.</strong></p><p><strong>3 The Evolution of the Software Engineer</strong></p><p>Historically, engineers were responsible for implementation. Tomorrow’s engineers may spend less time implementing and more time directing.</p><p>The progression might look like this:</p><p><strong>2023: Software Engineer: </strong>Writes code.<br><strong>2025: AI-Assisted Engineer: </strong>Writes prompts and reviews AI-generated code.<br><strong>2026: Agent Orchestrator: </strong>Coordinates multiple AI agents, Defines specifications and reviews outcomes, loop …</p><h3><strong>2027+: AI Business Engineer:</strong></h3><p>Combines: Product management, Customer discovery, AI orchestration, Automation, Software architecture and Business strategy.</p><p>The value shifts from coding to decision-making. Several industry observers describe this transition as a move from vibe coding toward agentic engineering, where humans increasingly define objectives while AI systems execute implementation details.</p><p>In a world where everyone can build, the advantage shifts to those who know what should be built and why. Vibe coding was never the destination. It was the first step.</p><p><strong>The next chapter is not about generating more code. It is about generating more value as results.</strong></p><p><strong>4 References:</strong></p><p>[Youtube $13B Lovable Co-Founder: Never Launch Your AI App for Free](<a href="https://www.youtube.com/watch?v=KEAYPqQG2-M">https://www.youtube.com/watch?v=KEAYPqQG2-M</a>)</p><p>[Harness engineering: leveraging Codex in an agent-first world](<a href="https://openai.com/index/harness-engineering/">https://openai.com/index/harness-engineering/</a>)</p><p>[Introduction to vibe coding](<a href="https://learn.microsoft.com/en-us/training/modules/introduction-vibe-coding/">https://learn.microsoft.com/en-us/training/modules/introduction-vibe-coding/</a>)</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=86512c792506" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[LangChain integrations with Ollama local LLM - qwen3.8:27b]]></title>
            <link>https://medium.com/@shmilysyg/langchain-integrations-with-ollama-local-llm-qwen3-8-27b-6b7a2c458611?source=rss-9bd333c8dcb5------2</link>
            <guid isPermaLink="false">https://medium.com/p/6b7a2c458611</guid>
            <category><![CDATA[ollama]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[langchain]]></category>
            <dc:creator><![CDATA[CellCS]]></dc:creator>
            <pubDate>Tue, 25 Aug 2026 04:33:05 GMT</pubDate>
            <atom:updated>2026-08-25T04:33:05.334Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/310/1*Xy817DbniirN2u46PuKIsQ.png" /></figure><p>Run open-source models (qwen3.8:27b) locally and wire them into langchain chains, agents, and tools — no API keys, no cloud, no per-token billing.</p><p><strong>1 Introduction:</strong></p><p><a href="https://ollama.com/"><strong>1.1 Ollama</strong></a>: Large language models don’t have to live behind a paid API. Ollama bundles model weights, configuration, and runtime into a single self-contained package defined by a Modelfile, and it handles the messy setup details — including GPU acceleration — for you.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*WKPaiARkgaS3O_71wk03pw.png" /><figcaption><a href="https://ollama.com/">https://ollama.com/</a></figcaption></figure><p>1.2 <strong>LangChain</strong> ships a first-class integration for Ollama through the <a href="https://docs.langchain.com/oss/python/integrations/chat/ollama">langchain-ollama package</a>, exposing everything you’d expect from a modern chat model: tool calling, structured output, image (multimodal) input, token-level streaming, async, and log probabilities, like <a href="https://ollama.com/library/qwen3.8:27b"><strong>qwen3.8:27b</strong></a>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/974/1*NktaXjKu9XKYC2bRekCnSQ.png" /><figcaption><a href="https://www.langchain.com/langchain">https://www.langchain.com/langchain</a></figcaption></figure><p><strong>2. Ollama Installation</strong></p><p>Follow <a href="https://ollama.com/download">https://ollama.com/download</a> to install Ollama. In this post, we use macOS as demo.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*H6Y-e49Xpi96MhKnvJTCOw.png" /><figcaption><a href="https://ollama.com/download">https://ollama.com/download</a></figcaption></figure><p><strong>3 Pull models into local</strong></p><p>Ollama download a model, here is using <strong><em>qwen3.8:27b</em></strong></p><pre>ollama pull qwen3.8:27b</pre><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*0wW0HH_I37aXxki71-Q5fA.png" /><figcaption>ollama pull qwen3.8:27b model</figcaption></figure><p>Check local models:</p><pre>ollama list</pre><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*i73wVob0kJVTr1heQE-OBw.png" /><figcaption>Local ollama models</figcaption></figure><p><strong>4 Ollama Chat Integration</strong></p><p>Download related <a href="https://github.com/CellCS/medium-ai-rnd/tree/main/langchain-playground">playground repo</a>. This repo is using uv and python 3.13. When you run uv sync, the tool automatically completes the entire setup process for you.</p><p><strong>4.1 Basic invocation — the base agent</strong></p><p><a href="https://github.com/CellCS/medium-ai-rnd/blob/main/langchain-playground/base-agent.py">base-agent.py</a> is the classic “hello world”: a system prompt plus a human message, translated in one call.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*iELfZ2bqJ6GLoGnMh3gOPg.png" /><figcaption><strong>base agent</strong></figcaption></figure><p>Instantiate ChatOllama just like any other LangChain chat model:</p><pre>from langchain_ollama import ChatOllama<br><br>llm = ChatOllama(<br>    model=&quot;qwen3.8:27b&quot;,<br>    temperature=0.3,<br>)</pre><p>Because ChatOllama implements LangChain’s BaseChatModel interface, it slots into invoke, stream, bind_tools, and chains exactly like GPT or Claude would.</p><p>Execute uv run base-agent.py</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*hptBDd_Gq0C3jipNyfoDOw.png" /><figcaption>uv run base-agent.py</figcaption></figure><p><strong>4.2 Tool calling</strong></p><p>Ollama’s tool support follows the OpenAI-compatible server spec, so you can use LangChain’s standard bind_tools() workflow. Define a tool with the <a href="http://twitter.com/tool">@tool</a> decorator and bind it to the model: (<a href="https://github.com/CellCS/medium-ai-rnd/blob/main/langchain-playground/toolcall.py">demo py file</a>)</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*8yFFTfk_e_oZl4BDQR_yZg.png" /><figcaption>Tool call in Langchain</figcaption></figure><p>execute uv run toolcall.py</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*O9GkDliSE4prqWvC67C4aA.png" /><figcaption>uv run toolcall.py</figcaption></figure><p>The model returns a structured tool_calls payload rather than free text — the foundation of LangChain agents.</p><p><strong>4.3 Reasoning models &amp; custom message roles</strong></p><p>Some models expose their “thinking” process through a non-standard control message role. Since control isn’t a standard role, we express it with a ChatMessage in <a href="https://github.com/CellCS/medium-ai-rnd/blob/main/langchain-playground/thinking.py">demo py file</a>:</p><pre>messages = [<br>    ChatMessage(role=&quot;control&quot;, content=&quot;thinking&quot;),<br>    HumanMessage(&quot;What is 3^3?&quot;),<br>]</pre><p><strong>Custom Message Roles</strong>: System Message vs. Prompt Text Structure</p><p>Use a dedicated <strong>SystemMessage</strong> to define the model&#39;s behavior, while put instructions directly inside the <strong>text content of a </strong><strong>HumanMessage</strong>.</p><p><strong>SystemMessage</strong>: Define the model’s behavior and is intended for high-level instructions, rules, and behavioral constraints. (<a href="https://github.com/CellCS/medium-ai-rnd/blob/main/langchain-playground/multi-modal-sysmsg.py">demo py file</a>)</p><pre>messages = [<br>    SystemMessage(<br>        content=&quot;&quot;&quot;<br>You are an expert at reading screenshots and technical documentation.<br><br>Rules:<br>- Extract information only from the image.<br>- Do not hallucinate missing values.<br>- Return concise answers.<br>- Preserve commands exactly as shown.<br>&quot;&quot;&quot;<br>    ),<br>    HumanMessage(<br>        content=[<br>            {<br>                &quot;type&quot;: &quot;image_url&quot;,<br>                &quot;image_url&quot;: f&quot;data:image/png;base64,{image_b64}&quot;,<br>            },<br>            {<br>                &quot;type&quot;: &quot;text&quot;,<br>                &quot;text&quot;: &quot;What is the download command for the Gemma 7B model?&quot;,<br>            },<br>        ],<br>    ),<br>]</pre><p>The model is being told “This is your role, and these are the rules you should follow.”</p><p>The <strong>HumanMessage</strong> then provides the actual task.This separation becomes especially useful when you have multiple questions that should all follow the same rules (The same system-level behavior can apply to all of those questions):</p><pre>SystemMessage<br>    ↓<br>&quot;You are a technical documentation analyst.<br> Extract information only from the image.<br> Do not guess.<br> Return concise answers.&quot;<br><br>HumanMessage #1<br>    ↓<br>&quot;What is the download command?&quot;<br><br>HumanMessage #2<br>    ↓<br>&quot;What version of the model is shown?&quot;<br><br>HumanMessage #3<br>    ↓<br>&quot;What is the model&#39;s parameter count?&quot;</pre><p>The text portion of a HumanMessage is better suited for task-specific instructions, like in <a href="https://github.com/CellCS/medium-ai-rnd/blob/main/langchain-playground/multi-modal-prompt-2.py">py file</a>:</p><pre>HumanMessage(<br>    content=[<br>        {<br>            &quot;type&quot;: &quot;image_url&quot;,<br>            &quot;image_url&quot;: f&quot;data:image/png;base64,{image_b64}&quot;,<br>        },<br>        {<br>            &quot;type&quot;: &quot;text&quot;,<br>            &quot;text&quot;: &quot;&quot;&quot;<br>The image is a screenshot from an Ollama model page.<br><br>Task:<br>1. Find the section containing model download instructions.<br>2. Locate the Gemma 7B model.<br>3. Return only the exact Ollama pull command.<br><br>Output format:<br>Return JSON only:<br><br>{<br>    &quot;model_name&quot;: &quot;&quot;,<br>    &quot;download_command&quot;: &quot;&quot;,<br>    &quot;confidence&quot;: &quot;&quot;<br>}<br><br>Do not include explanations.<br>&quot;&quot;&quot;<br>        },<br>    ]<br>)</pre><p>Here, everything is related to <strong>this particular image and this particular task</strong>. The prompt provides:</p><ul><li><strong>Image context</strong> — what the image represents</li><li><strong>Task</strong> — what the model needs to find</li><li><strong>Output requirements</strong> — how the answer should be formatted</li><li><strong>Restrictions</strong> — what the model should or should not return</li></ul><p>This is essentially a structured prompt.</p><p><strong>SystemMessage = behavior and policy</strong><br><strong>HumanMessage = task and input</strong></p><p>So put reusable behavioral instructions in SystemMessage; put request-specific context, questions, and output requirements in the HumanMessage text.</p><p><strong>4.4 Multi-modal image prompts</strong></p><p>Ollama supports vision models like qwen3.8:27b, gemma3 and bakllava. First convert your image to Base64, then compose a prompt with an image_url part and a text part. Here’s the full chain version (<a href="https://github.com/CellCS/medium-ai-rnd/blob/main/langchain-playground/multi-modal-chain.py">multi-modal-chain.py</a>):</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Mb-bTJqMtYA549KYBbyxAg.png" /></figure><p>execute uv run multi-modal-chain.py</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*vMBzRp13bzA4CTocPcFBZg.png" /><figcaption>uv run multi-modal-chain.py</figcaption></figure><p>Image source:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*GE8A4CCEF3W7z4t6opzKDg.png" /></figure><p><strong>5 Discussion</strong></p><p>With langchain-ollama, your local qwen3.8:27b model becomes a drop-in LangChain chat model: tool calling, reasoning, logprobs, and multimodal prompts all work through the same familiar interface. It’s the fastest way to prototype agents and chains without touching the cloud — and everything stays on your machine.</p><p>Happy reading, Happy coding!</p><p><strong>6 References:</strong></p><p>[ChatOllama integration docs](<a href="https://docs.langchain.com/oss/python/integrations/chat/ollama">https://docs.langchain.com/oss/python/integrations/chat/ollama</a>)</p><p>[Ollama](<a href="https://ollama.com/">https://ollama.com/</a>)</p><p>[Langchain](<a href="https://www.langchain.com/langchain">https://www.langchain.com/langchain</a>)</p><p>[Ollama model library](<a href="https://ollama.com/search">https://ollama.com/search</a>)</p><p>[demo py codes](<a href="https://github.com/CellCS/medium-ai-rnd/tree/main/langchain-playground">https://github.com/CellCS/medium-ai-rnd/tree/main/langchain-playground</a>)</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6b7a2c458611" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Tiny Idea That Changed AI Forever: Deep Residual Learning Explained for Everyone]]></title>
            <link>https://medium.com/@shmilysyg/the-tiny-idea-that-changed-ai-forever-deep-residual-learning-explained-for-everyone-96456615dd9e?source=rss-9bd333c8dcb5------2</link>
            <guid isPermaLink="false">https://medium.com/p/96456615dd9e</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[short-story]]></category>
            <dc:creator><![CDATA[CellCS]]></dc:creator>
            <pubDate>Wed, 03 Jun 2026 03:15:34 GMT</pubDate>
            <atom:updated>2026-06-03T03:15:34.994Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*fFohK1levb64o_ZJo8Ft-w.png" /></figure><p>In today’s world, AI can recognize faces, describe images, drive cars, and even help write articles like this one. But few people realize that one surprisingly simple idea helped make much of modern AI possible.</p><p>That idea is called <strong>Deep Residual Learning</strong>, better known as <strong>ResNet</strong>. The story of ResNet isn’t just a technical breakthrough. It’s a fascinating chapter in the history of artificial intelligence — a story involving decades of persistence, famous competitions, and researchers who refused to give up on an unpopular idea.</p><p><strong>Before ChatGPT, There Was ImageNet</strong></p><p>Long before AI could generate essays or create images, researchers faced a simpler challenge:</p><p><strong>Can a computer recognize what’s in a picture?</strong></p><p>A dog.</p><p>A bicycle.</p><p>A banana.</p><p>A school bus.</p><p>Easy for humans. Surprisingly difficult for computers.</p><p>To solve this problem, researchers created a massive image database called <strong>ImageNet</strong>, containing millions of labeled pictures. Every year, teams worldwide competed in the ImageNet challenge, trying to build the best image-recognition system. Think of it as the Olympics of computer vision. Winning ImageNet became one of the most prestigious achievements in AI.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/750/1*xVuSIuVMmGzUvYT3YRdcwA.png" /></figure><p>Set up by data scientist Fei-Fei Li in 2006, the ImageNet database now contains more than 14 million annotated images.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/500/0*eOJzFlesrvcsnOk5.png" /><figcaption><a href="https://www.historyofdatascience.com/imagenet-a-pioneering-vision-for-computers/">ImageNet: A Pioneering Vision for Computers</a></figcaption></figure><p><strong>The Man Who Refused to Give Up on Neural Networks</strong></p><p>To understand ResNet, we need to go back even further. Today, Geoffrey Hinton is often called the “Godfather of Deep Learning.” But for much of his career, many researchers believed neural networks were a dead end.</p><p>During the 1980s, 1990s, and early 2000s, AI research was dominated by other approaches. Neural networks were often considered slow, impractical, and unlikely to succeed.</p><p>For decades, Geoffrey Hinton continued studying neural networks when relatively few people cared about them. Then everything changed.</p><p><strong>The Earthquake of 2012: AlexNet</strong></p><p>In 2012, Hinton and two of his students, and, entered the ImageNet competition with a deep neural network called <strong>AlexNet</strong>. The result shocked the AI community. AlexNet didn’t just win. It crushed the competition.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*ETpBeZ6TG9RIfX72" /><figcaption><a href="https://www.zdnet.com/article/alexnet-the-ai-model-that-started-it-all-released-in-source-code-form-for-all-to-download/">AlexNet, the AI model that started it all, released in source code form — for all to download</a></figcaption></figure><p>Its error rate was dramatically lower than every other entry. Suddenly, neural networks were no longer an academic curiosity — they were the future.</p><p>Many historians of AI consider this the moment the modern deep learning revolution truly began. Almost overnight, companies and universities around the world started investing heavily in deep learning. The AI boom we see today can trace much of its origin back to this moment.</p><p><strong>The Next Problem: Going Deeper</strong></p><p>After AlexNet’s success, researchers noticed something interesting.</p><p>Generally speaking:</p><p><strong>Deeper neural networks performed better.</strong></p><p>More layers meant the AI could learn more complex patterns.</p><p>So researchers tried making networks deeper and deeper.</p><p>10 layers.</p><p>20 layers.</p><p>50 layers.</p><p>100 layers … …</p><p>But something strange happened.</p><p>At a certain point, performance actually got worse. This wasn’t because the networks were too small. It wasn’t because computers were too slow. The networks became difficult to train. Imagine teaching a message to pass through 100 people in a long line.</p><p>By the time the message reaches the last person, much of the original information has been lost. Deep neural networks faced a similar problem. The deeper they became, the harder they were to train effectively.</p><p><strong>The Genius of ResNet</strong></p><p>In 2015, a team from Microsoft Research led by introduced a remarkably simple solution. Instead of forcing every layer to learn everything from scratch, they added what became known as <strong>skip connections</strong>. Think of climbing a mountain. Normally, you must walk every step.</p><p>ResNet added shortcuts. Information could skip certain layers and flow directly to later layers.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*vvQLTt3jWvpg--lW.jpg" /><figcaption><a href="https://www.youtube.com/watch?v=woEs7UCaITo">ResNet Deep Neural Network Architecture Explained — YouTube</a></figcaption></figure><p>This simple idea helped information travel through very deep networks without getting lost.</p><p>The breakthrough was called <strong>Deep Residual Learning</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/0*Quzxy1ndruzXbG-i.jpg" /><figcaption><em>Jian Sun: </em><a href="https://blogs.microsoft.com/ai/microsoft-researchers-win-imagenet-computer-vision-challenge/">Microsoft researchers win ImageNet computer vision challenge</a></figcaption></figure><p>The word “residual” simply means that layers only needed to learn the difference — the leftover improvement — instead of rebuilding everything from scratch.</p><p>A useful analogy: Imagine editing a document. You don’t rewrite the entire article every time. You only modify the parts that need improvement.</p><p><em>ResNet applies a similar philosophy. </em>Each layer focuses on improving what’s already there.</p><p><strong>The ImageNet Victory That Changed Everything Again</strong></p><p>The result was extraordinary. Microsoft’s ResNet reached <strong>152 layers</strong>, which was incredibly deep at the time. Even more impressive, it won the 2015 ImageNet competition with a top-5 error rate of only 3.57%.</p><p>For comparison:</p><ul><li>Traditional systems before the deep learning era: around 25–30% error</li><li>AlexNet (2012): about 15.3% error. This breakthrough kickstarted the modern deep learning boom by nearly halving the error rate of previous, non-deep-learning models</li><li>ResNet (2015): about 3.57% error. It achieved this through an ensemble of 152-layer “residual networks” that solved the degradation problem (where deeper networks previously became less accurate).</li></ul><p>The progress in just three years was astonishing.</p><p>Researchers had gone from “deep learning might work” to “deep learning is surpassing human performance on some image-recognition tasks.”</p><p><strong>Why ResNet Matters Today</strong></p><p>Here’s the remarkable part.</p><p>Many people have never heard of ResNet.</p><p>Yet its influence is everywhere.</p><p>The idea of residual connections became one of the foundational building blocks of modern AI.</p><p>Many later systems borrowed or adapted the concept.</p><p>Researchers discovered that skip connections made training deep networks dramatically easier. Soon they appeared across computer vision and beyond.</p><p>The lesson was surprisingly simple:</p><p>Sometimes progress doesn’t come from making things more complicated.</p><p>Sometimes it comes from creating a shortcut.</p><p><strong>A Pattern We Keep Seeing in AI History</strong></p><p>Looking back, there is a fascinating pattern. First, Geoffrey Hinton spent decades defending neural networks when few people believed in them. Then AlexNet proved deep learning could outperform traditional approaches. Next, ResNet solved a major obstacle that limited how deep neural networks could become.</p><p>Each breakthrough built on the previous one. No single invention created modern AI. Instead, progress came from many researchers adding one important idea after another. Like building a cathedral, each generation laid a few more stones.</p><p><strong>The Real Lesson</strong></p><p>The story of Deep Residual Learning is not just about computers.It’s about persistence. Geoffrey Hinton spent decades pursuing an idea many people dismissed. The AlexNet team showed that deep learning could work at scale. The ResNet team solved a problem that seemed fundamental. Together, they helped create the foundation on which today’s AI systems stand.</p><p>When people talk about ChatGPT, self-driving cars, medical AI, or image generation, they are seeing the visible part of the iceberg.</p><p>Hidden underneath are breakthroughs like Deep Residual Learning — small ideas that quietly changed the world.</p><p>And perhaps that’s the most interesting lesson of all: History is often shaped not by giant leaps, but by simple ideas that arrive at exactly the right moment.</p><p>Happy Reading, Happy Coding!</p><p>References:</p><ul><li><a href="https://arxiv.org/abs/1512.03385">Deep Residual Learning for Image Recognition</a></li><li><a href="https://www.microsoft.com/en-us/research/blog/microsoft-researchers-algorithm-sets-imagenet-challenge-milestone/">Microsoft Researchers&#39; Algorithm Sets ImageNet Challenge Milestone - Microsoft Research</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=96456615dd9e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Winning with Simplicity: From Coachella to the AI Era]]></title>
            <link>https://medium.com/@shmilysyg/winning-with-simplicity-from-coachella-to-the-ai-era-c5cdd7b17c84?source=rss-9bd333c8dcb5------2</link>
            <guid isPermaLink="false">https://medium.com/p/c5cdd7b17c84</guid>
            <category><![CDATA[philosophy]]></category>
            <category><![CDATA[pop-music]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[CellCS]]></dc:creator>
            <pubDate>Wed, 15 Apr 2026 21:47:33 GMT</pubDate>
            <atom:updated>2026-04-15T21:49:45.559Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Rydjtg6kDV7jnObISI57Xw.png" /></figure><p>In 2026, Justin Bieber’s return to the Coachella stage after three years was remarkable not for its spectacle, but for its simplicity. There were no elaborate dance routines, no dazzling visual effects, and no grand theatrical staging. Dressed casually, he even opened his laptop on stage, searched for his old YouTube videos, and sang along with his past self. The performance felt unpolished and spontaneous — yet deeply authentic. Despite its minimalist approach, it captured global attention and sparked widespread discussion. It served as a powerful reminder that simplicity, when grounded in genuine connection, can resonate more profoundly than complexity.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/634/0*YHuCJmCJ15zCjSe8.jpg" /><figcaption><a href="https://www.dailymail.co.uk/tvshowbiz/article-15728155/Real-reason-Justin-Bieber-play-old-music-paid-10million-skip-songs-laptop-headline-Coachella-set.html">https://www.dailymail.co.uk/tvshowbiz/article-15728155/Real-reason-Justin-Bieber-play-old-music-paid-10million-skip-songs-laptop-headline-Coachella-set.html</a></figcaption></figure><p>This moment brings to mind Dr. Richard Sutton’s seminal 2019 essay, <strong><em>The Bitter Lesson</em></strong>. Dr. Sutton argues that general-purpose methods driven by computation — such as search and learning — consistently outperform systems built on handcrafted human knowledge. In the age of artificial intelligence, this principle extends far beyond technology. Time and again, scalable and adaptable approaches surpass intricate, specialized designs.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*QdkLG0f1pnEEsdXV.jpg" /></figure><p>However, it is important to maintain a rational perspective. This comparison is best understood as a philosophical metaphor rather than a strict causal relationship. Bieber’s success is built upon his established influence and loyal global fan base, just as the triumph of general-purpose AI relies on vast computational power and massive datasets. Simplicity, <strong>in both cases, is not born of limitation but enabled by scale, maturity, and accumulated strength.</strong></p><p>Seen through this lens, Bieber’s understated performance becomes a cultural echo of Sutton’s insight. Rather than relying on elaborate choreography or sophisticated stagecraft, he leveraged something more fundamental: authenticity, memory, and connection. The result illustrates a broader truth relevant to both art and innovation — what endures is not always the most complex solution, but the one built upon simple, universal foundations.</p><p>As we navigate the AI era, this lesson feels increasingly relevant. True progress is not always achieved by adding layers of sophistication, but by trusting scalable systems that can learn, adapt, and grow. Complexity may impress in the moment, but simplicity — rooted in universality — often wins in the long run.</p><p>Learn, adapt, and grow. Happy reading, happy coding!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c5cdd7b17c84" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Build PubMed MCP Server with FastAPI, FastMCP, and Python]]></title>
            <link>https://medium.com/@shmilysyg/build-pubmed-mcp-server-with-fastapi-fastmcp-and-python-9b771c3e2e5b?source=rss-9bd333c8dcb5------2</link>
            <guid isPermaLink="false">https://medium.com/p/9b771c3e2e5b</guid>
            <category><![CDATA[mcp-server]]></category>
            <category><![CDATA[biomedical]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[research]]></category>
            <dc:creator><![CDATA[CellCS]]></dc:creator>
            <pubDate>Wed, 01 Apr 2026 02:21:35 GMT</pubDate>
            <atom:updated>2026-04-01T02:21:35.321Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*5hV2uWQZ_TJCwPaH.png" /><figcaption>MCP Architecture</figcaption></figure><p><a href="https://pubmed.ncbi.nlm.nih.gov/"><strong>PubMed</strong></a><strong> </strong>is one of the most important resources for biomedical and life science literature, but most AI clients still need a clean bridge to search papers, retrieve metadata, and read abstracts in a structured way.</p><p>PubMed already exposes the excellent <a href="https://www.ncbi.nlm.nih.gov/home/develop/api/"><strong>NCBI E-utilities API,</strong></a> but raw API integration is not always the best interface for MCP-enabled AI clients. Most coding agents and AI assistants work better when they interact with a small, predictable set of tools instead of dealing directly with multiple HTTP endpoints, XML payloads, and API-specific parameters.</p><p>This project, <a href="https://github.com/CellCS/pubmed-mcp">pubmed-mcp</a> (<a href="https://github.com/CellCS/pubmed-mcp">https://github.com/CellCS/pubmed-mcp</a>), provides that bridge using: Python 3.13, FastAPI, FastMCP, httpx, Pydantic Settings, Docker-first deployment.</p><p>The goal is simple: make PubMed paper discovery and article retrieval available through a reusable MCP server that is easy to run locally or host centrally for a team or individual researcher.</p><p>The current implementation exposes three focused MCP tools:</p><ul><li><strong><em>search_pubmed</em></strong>(query, max_results, retstart, sort): search PubMed and return normalized article metadata</li><li><strong><em>get_pubmed_article</em></strong>(pmid): retrieve detailed metadata plus abstract for a specific PubMed article</li><li><strong><em>get_pubmed_abstract</em></strong>(pmid): retrieve only the abstract text for a PMID</li></ul><p>This is intentionally small and practical. Instead of trying to mirror every PubMed capability, the project focuses on the operations AI clients need most often: <strong><em>literature search, article lookup by PMID, abstract retrieval for summarization or downstream analysis</em></strong>.</p><p><strong>NCBI-aware rate limiting: </strong>The PubMed client enforces a request-per-second limit using an async lock and a minimum interval between requests. That matters because NCBI E-utilities usage should be respectful and predictable, especially if the service is shared by multiple users or agents.</p><p><strong>Required email and optional API key: </strong>NCBI expects clients to identify themselves properly. This server requires PUBMED_EMAIL and supports PUBMED_API_KEY to improve throughput.</p><p><strong>Bearer token protection: </strong>The MCP endpoint can be protected with one or more Bearer tokens using MCP_BEARER_TOKENS. If tokens are not configured, authentication is disabled. That gives flexibility for both local development and controlled shared deployment.</p><p><strong>Setup and deploy: </strong>Create .env from .env.example and configure these values:<br>- PUBMED_EMAIL: required by NCBI<br>- PUBMED_API_KEY: optional, improves throughput<br>- PUBMED_EUTILS_LIMIT: request rate limit, default 10<br>- MCP_HTTP_PATH: MCP route path, default /pubmed-mcp<br>- MCP_BEARER_TOKENS: optional comma-separated Bearer tokens</p><pre>PUBMED_EMAIL=your_email@example.com<br>PUBMED_API_KEY=your_pubmed_api_key<br>PUBMED_EUTILS_LIMIT=10<br>MCP_HTTP_PATH=/pubmed-mcp<br>MCP_BEARER_TOKENS=your-mcp-token-1,your-mcp-token-2</pre><p>Then build and run the container: <em>sh deploy-container.sh</em></p><p><strong>MCP client configuration: </strong>Use the following examples to connect your MCP client to the running service.</p><p><strong>OpenCode:</strong></p><pre>&quot;pubmed-mcp&quot;: {<br>  &quot;type&quot;: &quot;remote&quot;,<br>  &quot;url&quot;: &quot;http://127.0.0.1:8001/pubmed-mcp&quot;,<br>  &quot;oauth&quot;: false,<br>  &quot;headers&quot;: {<br>    &quot;Authorization&quot;: &quot;Bearer your-mcp-token-1&quot;<br>  }<br>}</pre><p><strong>Cline:</strong></p><pre>&quot;pubmed-mcp&quot;: {<br>  &quot;autoApprove&quot;: [],<br>  &quot;disabled&quot;: false,<br>  &quot;timeout&quot;: 60,<br>  &quot;type&quot;: &quot;streamableHttp&quot;,<br>  &quot;url&quot;: &quot;http://127.0.0.1:8001/pubmed-mcp&quot;,<br>  &quot;headers&quot;: {<br>    &quot;Authorization&quot;: &quot;Bearer your-mcp-token-1&quot;<br>  }<br>}</pre><p>After that, you can ask your AI client to:<br>- search for papers on a topic<br>- fetch a paper by PMID<br>- retrieve an abstract for summarization or research notes</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*JfNz4XfYrKUZ3gu6UJbo7A.png" /><figcaption>Test</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*WQpMKw5Q6QaSKuR7Mmxsgg.png" /><figcaption>Results</figcaption></figure><p><strong>Security notes: </strong>Even for a relatively small MCP service, this repo includes a useful security scan workflow.</p><p>The <strong><em>security-scan.sh</em></strong> script runs:</p><ul><li><strong>pip-audit </strong>for dependency vulnerabilities</li><li><strong>bandit</strong> for static code scanning</li><li><strong>gitleaks</strong> for secret scanning</li><li><strong>trivy</strong> for container image scanning</li><li><strong>syft</strong> for SBOM generation</li></ul><p>The generated artifacts are stored under s<strong><em>ecurity-scan-results/</em></strong>, and the repo includes a deployment checklist report summarizing the scan status.<br>That is a strong addition for a project like this because MCP servers are often treated as quick prototypes, but once they become shared team infrastructure, security posture and deployment hygiene start to matter a lot more.</p><p><strong>Discussion: </strong>This project is a good example of what a practical MCP server should look like:</p><ul><li>small tool surface</li><li>clear FastAPI/FastMCP integration</li><li>focused domain logic</li><li>safe XML parsing</li><li>optional authentication</li><li>Docker-friendly deployment</li><li>basic security scanning built in</li></ul><p>What makes it especially useful is that it does not try to become a full biomedical research platform. It stays focused on a narrow, high-value integration point: making PubMed accessible to MCP-enabled AI clients in a clean and reusable way.<br>For local use, it is simple enough to start in minutes. For team use, it already has the structure needed for a shared internal service: environment-based configuration, token protection, health check endpoint, container packaging, and security scan artifacts.</p><p>If you are building AI workflows for scientific literature search, research assistants, healthcare knowledge retrieval, or biomedical summarization, this is a very practical pattern to follow.</p><p>Happy Reading, Happy Coding.</p><p><strong>References:</strong></p><p>1. <a href="https://github.com/CellCS/pubmed-mcp">pubmed-mcp GitHub Repo</a> <br>2. <a href="https://www.ncbi.nlm.nih.gov/books/NBK25500/">NCBI E-utilities Quick Start</a> <br>3. <a href="https://www.ncbi.nlm.nih.gov/books/NBK25497">A General Introduction to the E-utilities</a> <br>4. <a href="https://fastapi.tiangolo.com/">FastAPI</a><br>5. <a href="https://github.com/jlowin/fastmcp">FastMC</a>P</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9b771c3e2e5b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Build a Secure, Read-Only Azure DevOps MCP Server with FastAPI, FastMCP, and Python]]></title>
            <link>https://medium.com/@shmilysyg/build-a-secure-read-only-azure-devops-mcp-server-with-fastapi-fastmcp-and-python-e5d9a2a677da?source=rss-9bd333c8dcb5------2</link>
            <guid isPermaLink="false">https://medium.com/p/e5d9a2a677da</guid>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[mcps]]></category>
            <category><![CDATA[mcp-server]]></category>
            <category><![CDATA[azure-devops]]></category>
            <category><![CDATA[ai-coding-agent]]></category>
            <dc:creator><![CDATA[CellCS]]></dc:creator>
            <pubDate>Mon, 16 Mar 2026 02:54:46 GMT</pubDate>
            <atom:updated>2026-03-16T02:57:16.131Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*lyuE4jIIwVZBr7Jwf1f99w.png" /></figure><p><strong>Azure DevOps</strong> has powerful REST APIs, but most AI clients still need a clean bridge to query work items, pull requests, and project data in a reusable way.</p><p><strong>1 Why this project exist</strong>s</p><p><strong>Microsoft’s Azure DevOps MCP</strong> is an excellent full-platform implementation with a very broad toolset, but here, <a href="https://github.com/CellCS/azure-devops-mcp"><strong>current project</strong></a> is better suited for teams that prioritize a lightweight, security-first, and centrally managed deployment model: it stays intentionally <strong>read-only</strong>, exposes a smaller and more predictable tool surface for LLM reliability, <strong>runs as a shared Dockerized HTTP MCP service</strong> instead of per-user local setup, and is easier to operate under enterprise governance where least privilege, auditability, and controlled integration boundaries matter more than maximizing feature coverage.</p><p><strong>This project provides that bridge using:</strong><br>- Python 3.13<br>- FastAPI + FastMCP<br>- Requests + Pydantic Settings<br>- Docker-first deployment<br>The goal is simple: make Azure DevOps data safely available to MCP-enabled clients without exposing write operations.</p><p><strong>2. What it currently supports</strong></p><p>Currently, the MCP server currently includes tools for:<br>- <strong>get_projects_list</strong>: list organization projects<br>- <strong>get_work_item_content</strong>: retrieve a work item by ID (project-scoped)<br>- <strong>get_related_work_item_info</strong>: fetch relations and related items<br>- <strong>get_test_workitem_steps</strong>: decode Azure DevOps Test Case steps<br>- <strong>get_pull_request_content</strong>: retrieve PR details by ID (project-scoped)<br>- <strong>query_work_items_by_wiql</strong>: run WIQL queries by project</p><p><strong>3. Architecture</strong></p><p>The service exposes a single MCP HTTP endpoint and delegates data access to Azure DevOps REST APIs.<br><strong>Client flow:</strong><br>1. MCP client calls tool endpoint with Bearer token<br>2. Server validates token<br>3. Server calls Azure DevOps API using PAT<br>4. Server returns normalized JSON response</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*FMSZuqpSQz3FCLmiXdh_iA.png" /><figcaption>Arch Overview</figcaption></figure><p><strong>It stays intentionally read-only.</strong></p><p><strong>4. Security-first decisions</strong></p><p>While Azure DevOps REST APIs are powerful, most AI clients still need a clean, reusable bridge for project, work item, pull request, and query data. This service provides that bridge through focused MCP tools like project listing, work item retrieval, related work item expansion, test step decoding, pull request lookup, and WIQL-based querying. A key design improvement <strong>was removing fixed project environment settings </strong>and <strong>making every request explicitly project-scoped</strong>, which is far more practical for real multi-project organizations.<br><strong>Security was a first-class design requirement from day one</strong>: the service is intentionally read-only, uses dual-layer authentication (<strong>MCP_BEARER_TOKENS for MCP access</strong> and AZURE_PAT for Azure DevOps), validates required environment configuration, URL-encodes project references, and enforces defensive HTTP behavior with timeouts and raise_for_status().</p><p><strong>It also includes a security scan pipeline</strong> (pip-audit, bandit, gitleaks, trivy, syft) with improved reliability through command exit-code based pass/fail detection. The result is a lightweight, enterprise-friendly MCP service that prioritizes least privilege, auditability, and safe integration over feature bloat, as shown in following table:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*_mY8jWB7m54DSwxqg3chXw.png" /><figcaption>Deployment controls checklist</figcaption></figure><p><strong>5. Setup and deploy</strong></p><p>Create .env by using .env.example. Then open .env and configure these required values:</p><pre>- AZURE_ORG: your Azure DevOps organization name (for example, contoso)<br>- AZURE_PAT: your Azure DevOps Personal Access Token (use least-privilege, read-only scopes)<br>- MCP_BEARER_TOKENS: one or more bearer tokens used by MCP clients to authenticate (comma-separated if multiple)</pre><p>After that, build and run the service with Docker:</p><pre>sh deploy-container.sh</pre><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2AH8D87_ZpS-QdbU0wdJ4g.png" /><figcaption>Docker container</figcaption></figure><p><strong>6. Connection Json configuration</strong></p><p>Use the following MCP client configurations to connect to your running azure-devops-mcp service. Here are just examples, remember to</p><ul><li>Replace your-mcp-token-1 with one of the tokens configured in MCP_BEARER_TOKENS</li><li>If your service runs on another host/port or path, update the url accordingly</li><li>Keep these configs local to your machine or secure workspace settings (do not commit real tokens</li></ul><p>OpenCode example</p><pre>&quot;azure-devops-mcp&quot;: {<br>  &quot;type&quot;: &quot;remote&quot;,<br>  &quot;url&quot;: &quot;http://127.0.0.1:8000/azure-devops-mcp&quot;,<br>  &quot;oauth&quot;: false,<br>  &quot;headers&quot;: {<br>    &quot;Authorization&quot;: &quot;Bearer your-mcp-token-1&quot;<br>  }<br>}</pre><p>Cline example</p><pre>&quot;azure-devops-mcp&quot;: {<br>  &quot;autoApprove&quot;: [],<br>  &quot;disabled&quot;: false,<br>  &quot;timeout&quot;: 60,<br>  &quot;type&quot;: &quot;streamableHttp&quot;,<br>  &quot;url&quot;: &quot;http://127.0.0.1:8000/azure-devops-mcp&quot;,<br>  &quot;headers&quot;: {<br>    &quot;Authorization&quot;: &quot;Bearer your-mcp-token-1&quot;<br>  }<br>}</pre><p>Then test in your coding agent tool.</p><p><strong>7 Discussion</strong></p><p>This project is intentionally designed as a lightweight Azure DevOps MCP service: a focused, security first, read-only toolset, minimal operational footprint, and Docker-friendly deployment make it easy to run and scale without the overhead of a full-platform integration. By limiting scope to high-value read operations and project-scoped requests, the service stays predictable for AI agents and easier to govern for engineering teams.</p><p>From a security and audit perspective, the architecture is practical for team and enterprise adoption. It includes layered authentication, explicit environment validation, reproducible security scanning, and artifact-based reporting to support compliance and traceability. Because of this balance between simplicity, security controls, and operational clarity, the service can be deployed as a shared internal gateway for a whole team, department, or larger enterprise environment where centralized access control is required.</p><p>Happy Reading, Happy Coding.</p><p><strong>8. References:</strong></p><ol><li><a href="https://github.com/CellCS/azure-devops-mcp"><strong>CellCS azure-devops-mcp</strong></a></li><li><a href="https://learn.microsoft.com/en-us/rest/api/azure/devops/wit/?view=azure-devops-rest-7.1"><strong>Microsoft Azure DevOps REST API</strong></a></li><li><a href="https://github.com/microsoft/azure-devops-mcp"><strong>Microsoft azure-devops-mcp</strong></a></li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e5d9a2a677da" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Agentic Coding in 2026: The Year Developers Became Directors — and Software Got Repriced]]></title>
            <link>https://medium.com/@shmilysyg/agentic-coding-in-2026-the-year-developers-became-directors-and-software-got-repriced-b38f6f2a1c05?source=rss-9bd333c8dcb5------2</link>
            <guid isPermaLink="false">https://medium.com/p/b38f6f2a1c05</guid>
            <category><![CDATA[coding]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[ai-agent]]></category>
            <dc:creator><![CDATA[CellCS]]></dc:creator>
            <pubDate>Wed, 25 Feb 2026 04:14:41 GMT</pubDate>
            <atom:updated>2026-02-25T04:14:41.712Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/546/1*P3T6uzpVtsggz8-t56RLEw.png" /><figcaption>Walking To The Sky (Captured in Summer 2025)</figcaption></figure><p><strong>Are software engineers busier after AI — or just busier managing it?</strong></p><p>Since coding agents became commercial-grade and easy to access, experimentation has exploded. Teams aren’t debating whether to use AI anymore. They’re testing multiple agents, orchestrating workflows, comparing output quality, and redesigning development cycles around them. Something fundamental has shifted. And it’s not just how we build software — it’s how software is valued.</p><p>Teams are no longer debating whether to use AI. That debate is over. Instead, they are:</p><ul><li>Testing multiple agents side by side</li><li>Designing workflows around them</li><li>Comparing output quality across models</li><li>Integrating them into CI/CD pipelines</li><li>Designing workflows around them</li><li>Revisiting development cycles to accommodate agent collaboration</li><li>Tools, MCPs, Skills, A2A …</li><li>(Tokens are costy, but worth? 😅)</li></ul><p>Developers are spending less time typing line-by-line implementations and more time supervising, validating, and orchestrating machine-generated output. The bottleneck is no longer syntax — it’s judgment.</p><p>Something fundamental has changed. Not just how software is built, but what we consider valuable engineering work. And this is precisely the shift documented in Anthropic’s 2026 Agentic Coding Trends Report.</p><p><strong>Inside the 2026 Agentic Coding Trends Report</strong></p><p>In early 2026, <strong>Anthropic</strong> published the <a href="https://resources.anthropic.com/hubfs/2026%20Agentic%20Coding%20Trends%20Report.pdf">2026 Agentic Coding Trends Report </a>— the first broad, empirical study of how modern developers actually use AI coding agents in real development environments.</p><p>The report surfaces how teams are integrating these systems into lived software workflows. It isn’t about whether AI can write code — it’s about how developers are choosing to delegate, orchestrate, and govern agentic behavior.</p><figure><img alt="" src="https://cdn-images-1.medium.com/proxy/1*uouMsWpiIXyzq2Xg-udZxg.png" /><figcaption>From <a href="https://resources.anthropic.com/hubfs/2026%20Agentic%20Coding%20Trends%20Report.pdf">https://resources.anthropic.com/hubfs/2026%20Agentic%20Coding%20Trends%20Report.pdf</a></figcaption></figure><p>1. Agentic Coding Is Expanding Across the SDLC</p><p>Developers aren’t just using AI for autocomplete or test generation anymore. Instead, agents are being deployed across the <strong>entire software development lifecycle</strong> — from design and refactoring to deployment and observability.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ba2hiqsnf8RUSQMmJ_obww.png" /><figcaption>From <a href="https://resources.anthropic.com/hubfs/2026%20Agentic%20Coding%20Trends%20Report.pdf">https://resources.anthropic.com/hubfs/2026%20Agentic%20Coding%20Trends%20Report.pdf</a></figcaption></figure><p>This mirrors what we already see in engineering teams: AI isn’t a helper tool anymore; it’s becoming a <em>participatory worker</em> that can act on a sequence of tasks or constraints, autonomously or semi-autonomously.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*S0bkH70qipZfEOioD8ar-w.png" /><figcaption>From <a href="https://resources.anthropic.com/hubfs/2026%20Agentic%20Coding%20Trends%20Report.pdf">https://resources.anthropic.com/hubfs/2026%20Agentic%20Coding%20Trends%20Report.pdf</a></figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xLfT4Q1IjrGX7GJqGcYMxg.png" /><figcaption>From <a href="https://resources.anthropic.com/hubfs/2026%20Agentic%20Coding%20Trends%20Report.pdf">https://resources.anthropic.com/hubfs/2026%20Agentic%20Coding%20Trends%20Report.pdf</a></figcaption></figure><p>2. Delegation Is High, but Full Autonomy Is Still Rare</p><p>One of the most important findings is a <em>spectrum of delegation. </em>While AI is used in a majority of tasks, developers only fully delegate a small portion of work to agents. Most of the time, humans remain in the loop — not because AI is bad, but because:</p><ul><li>Engineering work requires judgment</li><li>Edge cases are unpredictable</li><li>Security and correctness are non-negotiable</li><li>Need humans’ oversight and direction</li><li>Other factors …</li></ul><p>In practical terms this means agents are <strong>partners, not replacements</strong>. You don’t hand over the wheel — you share it.</p><p>3. Multi-Agent Workflows Are Emerging</p><p>Composing standalone models isn’t enough. What’s emerging are <strong>multi-agent workflows</strong> — systems where multiple specialized agents handle different parts of a larger task and coordinate between them.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*CNvq7HhIMvvDYRfq_U8T8g.png" /><figcaption>From <a href="https://resources.anthropic.com/hubfs/2026%20Agentic%20Coding%20Trends%20Report.pdf">https://resources.anthropic.com/hubfs/2026%20Agentic%20Coding%20Trends%20Report.pdf</a></figcaption></figure><p>Consider an example:</p><ul><li>One agent writes iterative tests</li><li>Another suggests code improvements based on test outcomes</li><li>A third drafts documentation</li><li>A fourth integrates the changes and checks deployment readiness</li></ul><p>This choreography transforms how engineering teams think about <em>who does what</em> — agents become collaborators in their own right. Human oversight shifts from reviewing everything to reviewing what matters.</p><p>4. A Shift Toward Long-Running Autonomous Execution</p><p>Another pattern identified is the rise of <strong>long-running agent execution</strong>, where agents are given a <em>goal</em> and allowed to work through it over extended periods, rather than in isolated synchronous prompts.</p><p>This is a meaningful departure from standard prompt-response usage, and it’s core to how agents are actually being used in production settings.</p><p>It moves the paradigm from reactive to proactive: the agent is not just responding to commands — it’s executing goals.</p><p>5. Agentic Coding Is Spreading <strong>Beyond Engineering Teams</strong></p><p>Perhaps the most revolutionary observation is that agentic coding is no longer isolated in engineering squads.</p><p><em>“In 2026, agentic coding is poised to expand into contexts and use cases that traditional development tools could not reach, from legacy languages to new form factors that democratize access beyond traditional developers.”</em></p><p>Teams in product, operations, data science, and even non-technical （like businesses …）functions are experimenting with agents to:</p><ul><li>Language barriers disappear (like COBOL recently …)</li><li>Generate internal tooling</li><li>Automate repetitive workflows</li><li>Assist with research and analysis</li><li>Synthesize insights from complex data</li></ul><p>This aligns with the broader shift we discussed above: AI tools are no longer <em>optional extras.</em> They’ve become <strong>ubiquitous utilities</strong> embedded in how organizations build and run software.</p><p><em>“We anticipate that one of the most significant trends in 2026 will be steady growth in agentic coding used by functional and business-process teams to create their own solutions to problems they experience, and improvements to processes they use every day.”</em></p><p>6. Productivity Gains Reshape Software Development Economics</p><p>Agentic coding doesn’t just speed up tasks — it changes the leverage of engineering work. When developers can supervise autonomous workflows that handle refactoring, testing, documentation, and modernization, output increases without proportional increases in headcount.</p><p>This shifts the bottleneck. The constraint is no longer typing speed or even raw implementation capacity. It becomes judgment, orchestration, and oversight. As a result:</p><ul><li>Engineering throughput expands</li><li>Legacy modernization becomes less costly</li><li>Iteration cycles compress</li><li>The value of pure implementation declines relative to system design and governance</li></ul><p>If productivity scales while human control remains central, software is not simply built faster. It is valued differently.</p><p>7. Agentic Coding Improves Security Defenses — But Also Expands Offensive Capabilities</p><p>The report highlights a dual reality: agentic systems strengthen defensive security workflows while simultaneously lowering barriers for offensive use. On the defensive side, coding agents can:</p><ul><li>Automatically identify vulnerabilities across large codebase</li><li>Generate patches and remediation suggestions</li><li>Expand test coverage and static analysis</li><li>Accelerate incident response and log review</li></ul><p>This increases the speed at which organizations can detect and respond to security issues — especially in complex or legacy systems.</p><p>But the same capabilities introduce risk. Agents capable of analyzing, generating, and modifying code can also be used to:</p><ul><li>Identify exploitable weaknesses</li><li>Generate attack scripts</li><li>Automate reconnaissance or vulnerability probing</li><li>Others …</li></ul><p>The report does not frame this as dystopian inevitability. Instead, it emphasizes governance. Agentic systems operate at scale, and scale amplifies both benefit and misuse.</p><p>Security, therefore, becomes less about preventing AI usage and more about controlling its deployment: defining guardrails, audit trails, access boundaries, and human approval loops. Agentic coding does not eliminate security challenges. It accelerates them — on both sides.</p><p>Conclusion and Discussion:</p><p><em>“The research is clear: AI is a constant collaborator, but using it effectively requires active supervision and validation, especially in high-stakes work. While more routine coding tasks can be delegated to AI, humans are still reviewing the code. It’s not “fully delegated” but highly collaborative. This distinction matters for how organizations approach AI adoption and how they think about the evolving role of engineers”</em></p><p><strong>Master Multi-Agent Coordination</strong>: The days of relying on a single “AI helper” are over. The real gains come from building <em>coordinated workflows</em> where multiple agents — each with a specialized role — work in parallel and hand tasks off reliably. This is about designing systems that break down complex problems and orchestrate agent collaboration at scale, not just calling an API.</p><p><strong>Scale Human-Agent Oversight</strong>: AI output will grow faster than humans can review it. The report stresses investing in <strong>automated pipeline review</strong>, where AI does preliminary validation and humans focus only on high-stakes decisions. This lets you direct attention where it matters most without becoming overwhelmed by noise.</p><p><strong>Extend Agentic Coding Beyond Engineering</strong>: Agentic coding isn’t just for developers anymore. Anthropic points to expanding these capabilities into non-engineering domains (e.g., legal, product, marketing), empowering domain experts to build tools and automation <em>without submitting tickets</em> to engineering teams. This matches real trends where organizations democratize AI tooling outside of software teams.</p><p><strong>Embed Security from Day One: Finally</strong> — and critically — the report highlights security as a <strong>foundational architectural priority</strong> rather than a post-hoc concern. With agents operating across production systems, security design must be built in from the start, both to protect against accidental vulnerabilities and to guard against potential misuse.</p><p>Happy Reading, Happy Coding! (may “Happy Oversight”)</p><h3><strong>References:</strong></h3><p>[anthropic: 2026 Agentic Coding Trends Report](<a href="https://resources.anthropic.com/hubfs/2026%20Agentic%20Coding%20Trends%20Report.pdf">https://resources.anthropic.com/hubfs/2026%20Agentic%20Coding%20Trends%20Report.pdf</a>)</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b38f6f2a1c05" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[2025 AI Summer Holiday Reading List: A Selection of Thought-Provoking Broadcasts and Publications]]></title>
            <link>https://medium.com/@shmilysyg/2025-ai-summer-holiday-reading-list-a-selection-of-thought-provoking-broadcasts-and-publications-358cd51fa7fd?source=rss-9bd333c8dcb5------2</link>
            <guid isPermaLink="false">https://medium.com/p/358cd51fa7fd</guid>
            <category><![CDATA[startup]]></category>
            <category><![CDATA[genai]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[CellCS]]></dc:creator>
            <pubDate>Wed, 16 Jul 2025 21:05:46 GMT</pubDate>
            <atom:updated>2025-07-16T21:05:46.178Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*05RSu3pVnjhq_0rbQdeEKw.png" /></figure><p>This summer’s AI landscape is marked by accelerating breakthroughs and ethical crossroads. Whether you’re lounging beachside or enjoying a quiet afternoon at home, these thought-provoking broadcasts and publications delve into AI’s implications — from practical survival strategies and societal impacts to advanced safety frameworks and the looming “intelligence explosion.” Each resource offers a window into how AI is reshaping our world and how we might shape AI in return.</p><p><strong>Trends — </strong><a href="http://: https://www.bondcap.com/report/pdf/Trends_Artificial_Intelligence.pdf"><strong>Artificial Intelligence (AI) — BOND</strong></a>: likely summarizes current and emerging trends in the field of AI. Key areas include the increasing use of AI for climate change solutions, the rise of human-centered AI that focuses on collaboration between humans and machines, and the growing impact of AI across various industries. Furthermore, the document may delve into specific AI techniques like deep learning, functional applications such as natural language processing and computer vision, and the diverse application fields where AI is being implemented</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/544/1*oYYx1eCV5Er7hnjsBRE_hw.png" /><figcaption>From: <a href="https://www.bondcap.com/report/pdf/Trends_Artificial_Intelligence.pdf">https://www.bondcap.com/report/pdf/Trends_Artificial_Intelligence.pdf</a></figcaption></figure><p><strong>Sonya Huang and Sarah Guo on what’s next in AI | Mercury Sphere</strong> : unpack the current state of AI investing. They explore which technologies are attracting the most capital and how startups across industries are integrating AI to stay ahead.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FtNL14eA8YSY%3Fstart%3D19%26feature%3Doembed%26start%3D19&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DtNL14eA8YSY&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FtNL14eA8YSY%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/3dfda4aa56e44fbd5a11d44f01a21960/href">https://medium.com/media/3dfda4aa56e44fbd5a11d44f01a21960/href</a></iframe><p><a href="https://www.youtube.com/watch?v=v9JBMnxuPX8"><strong>AI’s Trillion-Dollar Opportunity: Sequoia AI Ascent 2025 Keynote</strong></a></p><p>Sequoia Capital partners outline why AI represents a market opportunity at least 10x larger than cloud computing, where startups should focus to win, and how the rise of AI agents will create an entirely new economic paradigm. Founders need to adopt a “stochastic mindset” and “go at maximum velocity. All of the time.”</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2Fv9JBMnxuPX8%3Flist%3DPLOhHNjZItNnMEqGLRWkKjaMcdSJptkR08&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3Dv9JBMnxuPX8&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2Fv9JBMnxuPX8%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/4348cc58eceae492643cc0779f15179f/href">https://medium.com/media/4348cc58eceae492643cc0779f15179f/href</a></iframe><p>Kevin Scott, CTO @ Microsoft: <strong>An Evaluation of Deepseek and How We Underestimate the Chinese</strong></p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FKN7KYzpPfiU%3Fstart%3D587%26feature%3Doembed%26start%3D587&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DKN7KYzpPfiU&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FKN7KYzpPfiU%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/9f7edbb2b6312c3ecf6fc53c9416efb0/href">https://medium.com/media/9f7edbb2b6312c3ecf6fc53c9416efb0/href</a></iframe><p>OpenAI’s CPO on <strong>how AI changes must-have skills, moats, coding, startup playbooks, more | Kevin Weil</strong></p><p>Kevin Weil, OpenAI’s Chief Product Officer, shares insights on how OpenAI builds agile product teams, leverages model ensembles, and embraces “vibe coding” to drive innovation. He discusses the enduring power of chat interfaces, the growing importance of AI evals, and what traits OpenAI values in product managers — all while offering practical prompting tips for better AI use.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FscsW6_2SPC4%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DscsW6_2SPC4&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FscsW6_2SPC4%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/8f967ad9a6950d531a5f507464823ab1/href">https://medium.com/media/8f967ad9a6950d531a5f507464823ab1/href</a></iframe><p><strong>Andrew Ng: Building Faster with AI</strong></p><p>In this talk, Andrew Ng shares what he’s learning now: why execution speed matters more than ever, how agentic workflows are changing what startups can build, and why concreteness beats vagueness when turning ideas into products. He reflects on the rise of AI coding assistants, the shifting bottlenecks in product development, and why, despite faster software, it’s still human judgment and responsibility that will shape what comes next.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FRNJCfif1dPY%3Flist%3DPLQ-uHSnFig5NPx4adxl97CZb8vU4numwi&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DRNJCfif1dPY&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FRNJCfif1dPY%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/4af105adf5f839cea95d2464ea1052ef/href">https://medium.com/media/4af105adf5f839cea95d2464ea1052ef/href</a></iframe><p><strong>LLM</strong></p><p><strong>OpenAI: Pre-Training GPT-4.5</strong></p><p>Sam Altman sits down with Amin Tootoonchian, Alex Paino, and Daniel Selsam to discuss the journey to get to GPT-4.5.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2F6nJZopACRuQ%3Fstart%3D132%26feature%3Doembed%26start%3D132&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3D6nJZopACRuQ&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2F6nJZopACRuQ%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/41b4e9c7e156d66e5498a47d1490342c/href">https://medium.com/media/41b4e9c7e156d66e5498a47d1490342c/href</a></iframe><p><strong>AI Agents</strong></p><p>OpenAI: <a href="https://cdn.openai.com/business-guides-and-resources/ai-in-the-enterprise.pdf">AI in the Enterprise</a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/814/1*O_2R2Vd9yW__fTHoVGEVpQ.png" /><figcaption>From: <a href="https://cdn.openai.com/business-guides-and-resources/ai-in-the-enterprise.pdf">https://cdn.openai.com/business-guides-and-resources/ai-in-the-enterprise.pdf</a></figcaption></figure><p>OpenAI:<a href="https://cdn.openai.com/business-guides-and-resources/a-practical-guide-to-building-agents.pdf"> A practical guide to building agents</a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/982/1*cctfuHW3MvVI88OnQYVd0w.png" /><figcaption>From: <a href="https://cdn.openai.com/business-guides-and-resources/a-practical-guide-to-building-agents.pdf">https://cdn.openai.com/business-guides-and-resources/a-practical-guide-to-building-agents.pdf</a></figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/924/1*iIxNOBEE_AbWnyQq0CTXMA.png" /><figcaption>From: <a href="https://cdn-dynmedia-1.microsoft.com/is/content/microsoftcorp/microsoft/final/en-us/microsoft-brand/documents/Taxonomy-of-Failure-Mode-in-Agentic-AI-Systems-Whitepaper.pdf">https://cdn-dynmedia-1.microsoft.com/is/content/microsoftcorp/microsoft/final/en-us/microsoft-brand/documents/Taxonomy-of-Failure-Mode-in-Agentic-AI-Systems-Whitepaper.pdf</a></figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*EtfWYnfykxAo86iVzEKeMw.png" /><figcaption><em>Failure modes in agentic AI systems.(</em><a href="https://www.microsoft.com/en-us/security/blog/2025/04/24/new-whitepaper-outlines-the-taxonomy-of-failure-modes-in-ai-agents/">https://www.microsoft.com/en-us/security/blog/2025/04/24/new-whitepaper-outlines-the-taxonomy-of-failure-modes-in-ai-agents/</a><em>)</em></figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*i9WOKrOOkr1bvxx2rM6jmA.png" /><figcaption>From <a href="https://platform.openai.com/docs/guides/production-best-practices">https://platform.openai.com/docs/guides/production-best-practices</a></figcaption></figure><p>Claude Code: <a href="https://www.anthropic.com/engineering/claude-code-best-practices">Best practices for agentic coding</a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*8W3SWEM6dzprbbo0MQKE4A.png" /><figcaption>From <a href="https://www.anthropic.com/engineering/claude-code-best-practices">https://www.anthropic.com/engineering/claude-code-best-practices</a></figcaption></figure><p>Google: <a href="https://www.kaggle.com/whitepaper-agent-companion">Agents Companion</a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/762/1*G4ZpH_aD-7HdEmYj95c6aw.png" /><figcaption>From <a href="https://www.kaggle.com/whitepaper-agent-companion">https://www.kaggle.com/whitepaper-agent-companion</a></figcaption></figure><p><strong>Prompt</strong></p><p>Google: <a href="https://www.gptaiflow.tech/assets/files/2025-01-18-pdf-1-TechAI-Goolge-whitepaper_Prompt%20Engineering_v4-af36dcc7a49bb7269a58b1c9b89a8ae1.pdf">Prompt Engineering</a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qtsgIq1gIwm-BMdBA1qvog.png" /><figcaption>From: <a href="https://www.gptaiflow.tech/assets/files/2025-01-18-pdf-1-TechAI-Goolge-whitepaper_Prompt%20Engineering_v4-af36dcc7a49bb7269a58b1c9b89a8ae1.pdf">https://www.gptaiflow.tech/assets/files/2025-01-18-pdf-1-TechAI-Goolge-whitepaper_Prompt%20Engineering_v4-af36dcc7a49bb7269a58b1c9b89a8ae1.pdf</a></figcaption></figure><p>Google: <a href="https://services.google.com/fh/files/misc/gemini-for-google-workspace-prompting-guide-101.pdf">Prompting guild</a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Ut_rQje-F7YKP-aFam2PlQ.png" /><figcaption>From: <a href="https://services.google.com/fh/files/misc/gemini-for-google-workspace-prompting-guide-101.pdf">https://services.google.com/fh/files/misc/gemini-for-google-workspace-prompting-guide-101.pdf</a></figcaption></figure><p>Discussion: Together, these resources span the spectrum from foundational knowledge to forward-looking strategy. They invite a balanced summer reflection — equipping readers with ethical principles, practical understanding, data-backed insight, and the readiness to act thoughtfully in an era defined by AI.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=358cd51fa7fd" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[OpenAI Realtime API: Practical Guide with Real Case Example 1/2]]></title>
            <link>https://medium.com/@shmilysyg/openai-realtime-api-practical-guide-with-real-case-example-1-2-923e19afcaf2?source=rss-9bd333c8dcb5------2</link>
            <guid isPermaLink="false">https://medium.com/p/923e19afcaf2</guid>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[openai-realtime-api]]></category>
            <category><![CDATA[openai]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[CellCS]]></dc:creator>
            <pubDate>Mon, 28 Apr 2025 04:50:39 GMT</pubDate>
            <atom:updated>2025-04-28T04:50:39.240Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*zGk6l4gCnQl_sM-u13aq_A.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/969/1*yowdTzrynTAEIMP-x4Us5Q.png" /><figcaption>Realtime vs. STT-LLM-TTS</figcaption></figure><p>1 <strong>OpenAI Realtime API</strong></p><p>The OpenAI Realtime API enables low-latency, multimodal interactions including speech-to-speech conversational experiences and real-time transcription.</p><p>This API works with <strong>natively multimodal models</strong> like <a href="https://platform.openai.com/docs/models/gpt-4o-realtime"><strong>GPT-4</strong></a><strong>o</strong> and <a href="https://platform.openai.com/docs/models/gpt-4o-mini-realtime"><strong>GPT-4o mini</strong></a>, offering:</p><ul><li>Real-time <strong>text</strong> and <strong>audio</strong> processing.</li><li><strong>Function calling</strong> during conversations.</li><li><strong>Speech generation</strong> (assistant speaking back).</li><li><strong>Streaming transcription</strong> with models like <a href="https://platform.openai.com/docs/models/gpt-4o-transcribe"><strong>GPT-4o Transcribe</strong></a> and <a href="https://platform.openai.com/docs/models/gpt-4o-mini-transcribe"><strong>GPT-4o mini Transcribe</strong></a><strong>.</strong></li></ul><p>The Realtime API can also be used independently for transcription and turn detection use cases. For example, you can <strong>stream audio in</strong>, and receive <strong>live transcripts</strong> when speech is detected — without any chat experience.</p><p><strong>2 Use cases:</strong></p><ul><li>Build a <strong>real-time, speech-to-speech</strong> conversational AI.</li><li>Create <strong>live transcription</strong> applications.</li><li>Build <strong>voice agents</strong> (for phone calls, mobile apps, etc.).</li></ul><p><strong>3 Connect methods</strong>: There are two main ways to connect to the Realtime API.</p><ul><li>Using <a href="https://platform.openai.com/docs/guides/realtime#connect-with-webrtc">WebRTC</a>, which is ideal for client-side applications (for example, a web app)</li><li>Using <a href="https://platform.openai.com/docs/guides/realtime#connect-with-websockets">WebSockets</a>, which is great for server-to-server applications (from your backend or if you’re building a voice agent over phone for example), like browser and mobile clients.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*epKBJnNyG-NR201C" /><figcaption>From: <a href="https://platform.openai.com/docs/guides/realtime">https://platform.openai.com/docs/guides/realtime</a></figcaption></figure><p><strong>WebSocket</strong> is a communication protocol that enables full-duplex, real-time communication over a single, long-lived connection between a client (such as a web browser) and a server. Unlike the traditional HTTP request-response model, WebSocket allows the server to push data to the client at any time without waiting for a new request.<br>This makes it ideal for applications that require instant updates, such as online games, chat applications, live news feeds, or real-time speech interactions.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*74tTkrDk_0d6ALaW.png" /><figcaption>From:<a href="https://gcore.com/learning/what-is-websocket">https://gcore.com/learning/what-is-websocket</a></figcaption></figure><p>4. <strong>Response Lifecycle: Event Ordering</strong></p><p>The server might then send these intermediate events as it processes the <a href="https://platform.openai.com/docs/guides/realtime-conversations">response</a> in following <strong>lifecycle order:</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*EZoEkBWbKTMC87hf6MaZ5A.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*steq6yK15O1LYWsM1XoE3w.png" /><figcaption>From: <a href="https://platform.openai.com/docs/guides/realtime-conversations#handling-audio-with-websockets">https://platform.openai.com/docs/guides/realtime-conversations#handling-audio-with-websockets</a></figcaption></figure><p>Here is one realtime testing log for this speech-to-speech</p><pre>// in speech<br>User: &quot;Hello&quot;<br>Assistant: &quot;Hi there! Thanks for joining me today. I&#39;m exploring ….&quot;</pre><p>Workflow: “The <strong>user audio</strong> was streamed in.” -&gt; “The API generated a <strong>transcription</strong> for the user’s speech.” -&gt;Then the API streamed <strong>assistant audio</strong> (as raw PCM16 data) back to the client, along with <strong>transcripts</strong> and <strong>response markers</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*SJvRy8Qtnz4ds2UoxTfJwg.png" /><figcaption>One conversation log</figcaption></figure><p>Here’s a <strong>simplified overview</strong> of the event sequence that need handlers:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*uI9UTuiUMQGDNG4fmSiDqw.png" /></figure><p>Assistant Audio Streaming: Assistant’s reply audio is streamed <strong>chunk-by-chunk</strong> in response.audio.delta. Listen to response.audio.delta to buffer/play the assistant&#39;s audio stream. You need to <strong>assemble</strong> these chunks to play back smoothly. like convert pcm16 to wav, then send bytes to frontend client websocket.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ZRk5xjIxPrN5aGjBJGeViA.png" /><figcaption>event: “conversation.item.input_audio_transcription.completed”</figcaption></figure><p><em>“response.audio_transcript.done”</em>: Assistant’s whole transcripts. (Assistant’s transcripts are also included in events: “<em>response.audio_transcript.done”, “response.content_part.done”, “response.output_item.done”, “response.done”</em>). If you want to show <strong>“typing”-style live subtitles,</strong> then need to decode event <em>“response.audio_transcript.delta”.</em></p><p>“response.done”: indicate current <em>response.status:”completed”, also includes Assistant’s transcript. </em>Wait for response.done before accepting new user input.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*DnmQjBrn4iYeIbNNEPW74w.png" /><figcaption>event:response.done</figcaption></figure><p><strong>5 VS.</strong> Speech-to-Text (STT) and Text-to-Speech (TTS)</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*I42BfXx2ne-QZhAbHzBbhQ.png" /></figure><p>STT and TTS systems are typically separate services (Mic → STT → LLM → TTS → Speaker):</p><ul><li>STT converts speech → text (for example, a voice note transcriber).</li><li>LLM processing or AI Agents …</li><li>TTS converts text → speech (for example, a screen reader).</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/1*mixGtkCl3C6HncqmYexyJA.png" /><figcaption>From:<a href="https://platform.openai.com/docs/guides/audio">https://platform.openai.com/docs/guides/audio</a></figcaption></figure><p>In contrast, the OpenAI Realtime API offers a unified, real-time, multimodal experience (Mic → Realtime API → Speaker):</p><ul><li>It can transcribe the user’s speech live (like STT)</li><li>Understand the meaning</li><li>Generate a dynamic response (text and/or function calls)</li><li>And stream back synthesized speech in real-time (like TTS)</li><li>All happening within one continuous interaction.</li><li>(Tokens: approximately $0.06 per minute of audio input and $0.24 per minute of audio output.)</li></ul><p>Next, we will show real case like Flutter-web as frontend (client ws), FastAPI as backend with using web socket, OpenAI Realtime API (server ws).</p><p>Happy reading, happy coding.</p><p><strong>6 References:</strong></p><p>[OpenAI: Realtime API](<a href="https://platform.openai.com/docs/guides/realtime?connection-example=python&amp;use-case=conversation">https://platform.openai.com/docs/guides/realtime?connection-example=python&amp;use-case=conversation</a>)</p><p>[OpenAI:Realtime conversations](<a href="https://platform.openai.com/docs/guides/realtime-conversations">https://platform.openai.com/docs/guides/realtime-conversations</a>)</p><p>[OpenAI: Realtime — Websocket](<a href="https://platform.openai.com/docs/guides/realtime-conversations">https://platform.openai.com/docs/guides/realtime-conversations</a>)</p><p>[OpenAI: Realtime events](<a href="https://platform.openai.com/docs/api-reference/realtime-client-events">https://platform.openai.com/docs/api-reference/realtime-client-events</a>)</p><p>[OpenAI: Voice Agent](<a href="https://platform.openai.com/docs/guides/voice-agents">https://platform.openai.com/docs/guides/realtime?connection-example=python&amp;use-case=conversation</a>)</p><p>[OpeanAI: Agents SDK](<a href="https://openai.github.io/openai-agents-python/voice/quickstart/">https://openai.github.io/openai-agents-python/voice/quickstart/</a>)</p><p>[OpenAI: Audio and speech](<a href="https://platform.openai.com/docs/guides/audio">https://learn.microsoft.com/en-us/azure/ai-services/openai/how-to/realtime-audio#server-decision-mode</a>)</p><p>[OpenAI: Function calling](<a href="https://platform.openai.com/docs/guides/function-calling?api-mode=chat&amp;example=send-email">https://platform.openai.com/docs/guides/function-calling?api-mode=chat&amp;example=send-email</a>)</p><p>[Azure: How to use the GPT-4o Realtime API for speech and audio](<a href="https://learn.microsoft.com/en-us/azure/ai-services/openai/how-to/realtime-audio#server-decision-mode">https://learn.microsoft.com/en-us/azure/ai-services/openai/how-to/realtime-audio#server-decision-mode</a>)</p><p>[AG2: RealtimeAgent 2](<a href="https://docs.ag2.ai/0.8.5/docs/blog/2025/01/08/RealtimeAgent-over-websocket/">https://docs.ag2.ai/0.8.5/docs/blog/2025/01/08/RealtimeAgent-over-websocket/</a>)</p><p>[AG2: RealtimeAgent 1 ](<a href="https://docs.ag2.ai/docs/blog/2024/12/20/RealtimeAgent/">https://docs.ag2.ai/docs/blog/2024/12/20/RealtimeAgent/</a>)</p><p>[Flutter: record package](<a href="https://pub.dev/documentation/record/latest/">https://pub.dev/documentation/record/latest/</a>)</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=923e19afcaf2" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>