<?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 Poojan Vig on Medium]]></title>
        <description><![CDATA[Stories by Poojan Vig on Medium]]></description>
        <link>https://medium.com/@poojanvig.pv?source=rss-e432d5e4ef26------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*aPSqFUsvomX3w0KEXLJWFg.jpeg</url>
            <title>Stories by Poojan Vig on Medium</title>
            <link>https://medium.com/@poojanvig.pv?source=rss-e432d5e4ef26------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sat, 21 Nov 2026 20:14:27 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@poojanvig.pv/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[Claude Code Agent Teams: The End of Solo AI Coding?]]></title>
            <link>https://pub.towardsai.net/claude-code-agent-teams-the-end-of-solo-ai-coding-45da2cab6153?source=rss-e432d5e4ef26------2</link>
            <guid isPermaLink="false">https://medium.com/p/45da2cab6153</guid>
            <category><![CDATA[agents]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[claude-code]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[claude]]></category>
            <dc:creator><![CDATA[Poojan Vig]]></dc:creator>
            <pubDate>Sun, 15 Feb 2026 20:01:00 GMT</pubDate>
            <atom:updated>2026-02-15T20:01:00.819Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*jHjdjbxn8YabwZzCKtcmBw.jpeg" /></figure><p>Every developer who’s used Claude Code knows the drill — you give it a task, it spins up a sub-agent, the sub-agent works in complete isolation, finishes, and sends back a summary. It works, but it’s fundamentally limited. The sub-agent has no idea what other agents are doing. There’s no collaboration, no shared context, and no way for agents to challenge or build on each other’s work.</p><p>That just changed.</p><p>Claude Code just shipped <strong>Agent Teams</strong> — a completely new execution model where 3 to 5 independent Claude Code instances work together on the same project. They share context, exchange messages, coordinate through a shared task system, and even debate each other’s approaches like a real engineering team would.</p><p>This isn’t a minor upgrade to the old task and sub-agent model. It’s a fundamentally different way of thinking about how AI agents collaborate on complex coding problems.</p><p>In this blog, we’ll break down everything you need to know about Agent Teams — how to set them up, how they actually work under the hood, how they compare to the old sub-agent approach, and when you should (and shouldn’t) use them.</p><p>Let’s dive in.</p><h4>How to Install &amp; Enable Agent Teams</h4><p>Before you can start using Agent Teams, there are a few things you need to set up. It’s straightforward, but you need to get each step right.</p><p><strong>Step 1: Update Claude Code to the Latest Version<br></strong>Agent Teams are a recent addition, so make sure you’re running the latest version of Claude Code. If you’re not sure, just update it to be safe.</p><p><strong>Step 2: Enable the Experimental Feature Flag<br></strong>Agent Teams are still behind an experimental flag, which means they’re not turned on by default. You need to manually enable them.</p><p>Open your settings file by running this command in your terminal:</p><pre>code ~/.claude/settings.json</pre><p>Then add the following to your global settings:</p><pre>{<br>  &quot;env&quot;: {<br>    &quot;CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS&quot;: &quot;1&quot;<br>  }<br>}</pre><p>Save the file and restart your terminal. That’s it Agent Teams are now enabled.</p><p><strong>Step 3: Start a New Session and Trigger a Team<br></strong>Here’s something important to understand — Agent Teams don’t activate automatically. They kick in when your prompt explicitly asks Claude Code to create a team. <br>For example, you could write something like:</p><pre>&quot;I&#39;m building a CLI tool that tracks TODO comments across a codebase. <br>Create an agent team to explore this from multiple angles - <br>one teammate focused on UX, one on technical architecture, <br>and one playing devil&#39;s advocate.&quot;</pre><p>When Claude Code detects that intent, it starts creating team members on its own.</p><p><strong>The Best Way to Experience Agent Teams: Multi-Pane View.<br></strong>Agent Teams truly shine when you can watch every agent working in parallel. The best setup for this is using <strong>tmux</strong> or <strong>iTerm2</strong> on macOS. If you’re using iTerm2, here’s how to set it up: 1. Install iTerm2 if you haven’t already 2. Go to Settings → General → Magic 3. Enable the Python API <br>4. Restart iTerm2</p><p>Then launch Claude Code in teammate mode:</p><pre>claude --teammate-mode tmux</pre><p>This opens up separate panes one for the team lead and one for each teammate agent. You can click into any pane, watch what that specific agent is doing in real time, and even send direct messages to individual agents.</p><p>It feels less like running a tool and more like managing a small engineering team live in your terminal.</p><h4>Sub-Agents vs Agent Teams: What Actually Changed?</h4><p>To really appreciate what Agent Teams bring to the table, you need to understand how the old model worked and where it fell short.</p><p><strong>The Old Model: Sub-Agents</strong></p><p>Before Agent Teams, Claude Code followed a simple and linear execution pattern. The main agent would receive your task, spin up a sub-agent using the Task tool, and that sub-agent would go off and work in complete isolation. Once it finished, the session terminated and only a brief summary was sent back to the main agent.</p><p>That’s it. No communication between agents. No shared understanding of the bigger picture. If you had multiple sub-agents running, none of them knew the others existed. Each one was essentially working with blinders on.</p><p>This works fine for simple, self-contained tasks. But the moment you need agents to build on each other’s findings, challenge assumptions, or coordinate across different parts of a codebase, the old model breaks down fast.</p><p><strong>The New Model: Agent Teams</strong></p><p>Agent Teams flip this entirely. Instead of isolated sub-agents working in silos, you now get a team of 3 to 5 agents that operate with full awareness of each other. Here’s what’s different:</p><p><strong>Shared Task Lists</strong> — Every agent on the team can see what tasks exist, who’s working on what, and what’s still pending. There’s no duplication of effort and no blind spots.</p><p><strong>Direct Communication</strong> — Agents can send messages to each other. Not just summaries after the work is done, but real-time messages during execution. One agent can ask another for clarification, share a finding, or flag a conflict.</p><p><strong>Broadcast Messaging</strong> — An agent can send a message to every teammate at once. This is especially useful when one agent discovers something that changes the direction for the entire team.</p><p><strong>Lifecycle Control</strong> — The team lead agent can explicitly start and shut down teammates. There’s a proper startup and shutdown flow with confirmation, not just a process that runs and dies.</p><p>Think of it this way — sub-agents are like hiring five freelancers who never talk to each other and just email you their final reports. Agent Teams are like putting five engineers in the same room with a shared whiteboard and a project tracker.</p><h4>Inside Agent Teams: The 5 Core Tools Explained</h4><p>Now let’s get into the mechanics. Agent Teams are powered by five internal tools that handle everything from creating the team to shutting it down. Understanding these tools gives you a clear picture of how agents actually coordinate behind the scenes.</p><p><strong>Tool 1: TeamCreate</strong></p><p>Everything begins here. When Claude Code decides to form a team, it invokes the TeamCreate tool. This creates a new team folder under .claude/teams/ in your project directory. At this point, the team technically exists but no agents have been assigned yet. Think of it as setting up the workspace before anyone walks in — the scaffolding is ready, but the work hasn&#39;t started.</p><p><strong>Tool 2: TaskCreate</strong></p><p>This is where the actual work gets defined. The TaskCreate tool is specifically for creating tasks, not for spinning up agents. Each task is stored as a JSON file under .claude/tasks/team-id and tracks several key pieces of information — a unique task ID, a description of what needs to be done, the current status (pending, in progress, complete, or deleted), who owns it, and any dependencies between tasks using a blocks and blocked-by relationship.</p><p>Tasks can flow in two directions. The team lead can delegate them top-down, or individual agents can self-claim tasks by browsing the task list using tools like taskList or getTask.</p><p><strong>Tool 3: Task Tool (Upgraded)</strong></p><p>If you’ve used Claude Code before, you already know the Task tool — it’s what spins up sub-agent sessions. With Agent Teams, this same tool got a significant upgrade. It now accepts two new parameters: name and team_name. When you pass both of these, instead of launching an isolated sub-agent, it creates a full teammate agent that&#39;s connected to the team&#39;s shared context and communication system. Without those parameters, it still behaves like the old sub-agent model. Same tool, but a completely different execution path depending on how it&#39;s called.</p><p><strong>Tool 4: TaskUpdate</strong></p><p>Once agents are working, they need a way to report progress. The TaskUpdate tool handles this. Each agent is expected to call TaskUpdate to claim a task, mark it as in progress, flag it as complete, or update its status in any other way. This is what keeps the shared task list accurate and lets every teammate know where things stand at any given moment.</p><p><strong>Tool 5: SendMessage</strong></p><p>This is the tool that makes Agent Teams fundamentally different from sub-agents. The SendMessage tool enables real-time communication between agents in two ways — direct messages from one agent to another, and broadcast messages from one agent to all teammates.</p><p>Under the hood, messages are written to .claude/teams/&lt;team_id&gt;/inbox/, where each agent has its own inbox folder. When a message arrives, it gets injected into the receiving agent&#39;s conversation history as a new user message formatted like:</p><pre>&lt;teammate-message teammate_id=&quot;team-lead&quot;&gt;<br>  Your message content here<br>&lt;/teammate-message&gt;</pre><p>This means the receiving agent processes the message just like it would process any instruction, making the communication feel natural within the agent’s workflow.</p><p><strong>The Shutdown Flow</strong></p><p>When the work is done, the team lead sends a shutdown_request message to each teammate. The teammate responds with a shutdown_response to confirm, and a postToolCall hook auto-terminates that agent&#39;s session. It&#39;s a clean, controlled shutdown — no orphaned processes or hanging sessions.</p><p>The entire lifecycle, from team creation to shutdown, follows a clear pattern: <strong>Create the team → Define tasks → Spin up agents → Work and communicate → Update progress → Shut down.</strong> Every step is tracked and every interaction is logged.</p><h3>Real Use Case: Deep Debugging with Agent Teams</h3><p>The best way to understand when Agent Teams shine is through a real scenario. One example from Anthropic’s own documentation paints a clear picture.</p><p>Imagine users are reporting a bug — your app exits after a single message instead of staying connected. With the old sub-agent model, you’d describe the problem to Claude Code and it would investigate one hypothesis at a time, sequentially. If the first guess is wrong, it moves to the next. It’s slow, linear, and each attempt has no awareness of what was already tried.</p><p>With Agent Teams, the approach is completely different. You tell Claude Code to spawn five teammates, each investigating a different hypothesis simultaneously. One agent looks at the WebSocket connection handling. Another digs into the session timeout logic. A third examines the event loop. A fourth reviews the error handling middleware. And a fifth checks the client-side reconnection logic.</p><p>Here’s where it gets interesting — these agents don’t just work in parallel, they actively talk to each other. If one agent finds evidence that rules out another agent’s hypothesis, it sends a message saying so. They challenge each other’s theories like a scientific debate, narrowing down the root cause through collaboration rather than brute-force sequential investigation.</p><p>Once a consensus emerges, the findings are documented and the team shuts down cleanly.</p><p>This is something sub-agents simply cannot do. The combination of parallel execution, shared task visibility, and real-time messaging creates a debugging workflow that’s closer to how a real engineering team would tackle a complex bug — multiple people, multiple angles, constant communication.</p><h4>Trade-offs and Final Thoughts</h4><p>Agent Teams are impressive, but they’re not the right tool for every situation. There are real trade-offs you should know before reaching for them.</p><p>The biggest one is <strong>token consumption</strong>. When you’re running 3 to 5 agents simultaneously, each with its own conversation history and each exchanging messages, the token usage adds up fast. For simple or well-defined tasks where a single agent or a basic sub-agent can get the job done, using a full team is overkill and unnecessarily expensive.</p><p>The second trade-off is <strong>speed</strong>. More agents doesn’t always mean faster results. The coordination overhead — creating tasks, exchanging messages, waiting for responses — introduces latency that a single focused agent wouldn’t have. If your task is straightforward, a sub-agent will likely finish faster.</p><p>So when should you actually use Agent Teams? Use them when the problem is <strong>complex, multi-faceted, and benefits from multiple perspectives</strong>. Deep debugging, architectural exploration, large-scale refactoring across multiple modules, or any scenario where you’d normally want several engineers looking at different parts of the same problem — that’s where Agent Teams earn their cost.</p><p>For everything else, sub-agents still do the job just fine. It’s unlikely that Anthropic will sunset sub-agents anytime soon. Think of it as two tools in the same toolbox — one for precision tasks and one for collaborative problem-solving.</p><p>Agent Teams represent a genuine shift in how AI coding assistants work. We’ve gone from isolated agents that return summaries to collaborative teams that think, communicate, and build on each other’s ideas. As this feature matures and moves out of the experimental phase, the possibilities for long-running, multi-agent task completion are going to get very exciting.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=45da2cab6153" width="1" height="1" alt=""><hr><p><a href="https://pub.towardsai.net/claude-code-agent-teams-the-end-of-solo-ai-coding-45da2cab6153">Claude Code Agent Teams: The End of Solo AI Coding?</a> was originally published in <a href="https://pub.towardsai.net/">Towards AI</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[DeepSeek OCR — More that your OCR]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://pub.towardsai.net/deepseek-ocr-more-that-your-ocr-fa176238f523?source=rss-e432d5e4ef26------2"><img src="https://cdn-images-1.medium.com/max/1920/1*3ulkC8F6GWzIr4M9fAOzlg.png" width="1920"></a></p><p class="medium-feed-snippet">How AI researchers are using images to compress text 10&#xD7;&#x200A;&#x2014;&#x200A;and what it means for the future of AI memory</p><p class="medium-feed-link"><a href="https://pub.towardsai.net/deepseek-ocr-more-that-your-ocr-fa176238f523?source=rss-e432d5e4ef26------2">Continue reading on Towards AI »</a></p></div>]]></description>
            <link>https://pub.towardsai.net/deepseek-ocr-more-that-your-ocr-fa176238f523?source=rss-e432d5e4ef26------2</link>
            <guid isPermaLink="false">https://medium.com/p/fa176238f523</guid>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[deepseek]]></category>
            <category><![CDATA[programming]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Poojan Vig]]></dc:creator>
            <pubDate>Sat, 25 Oct 2025 14:01:57 GMT</pubDate>
            <atom:updated>2025-10-25T14:01:57.429Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Remember That Free Google CLI That Killed Cursor? It Just Got 10x Better]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://pub.towardsai.net/remember-that-free-google-cli-that-killed-cursor-it-just-got-10x-better-e5c878d4e14e?source=rss-e432d5e4ef26------2"><img src="https://cdn-images-1.medium.com/max/2435/1*XYYZgIRObV4BiPeo9J4gyQ.png" width="2435"></a></p><p class="medium-feed-snippet">Thousands of developers ditching Cursor and Windsurf overnight. Two months ago, I wrote about how Google&#x2019;s free Gemini CLI was disrupting&#x2026;</p><p class="medium-feed-link"><a href="https://pub.towardsai.net/remember-that-free-google-cli-that-killed-cursor-it-just-got-10x-better-e5c878d4e14e?source=rss-e432d5e4ef26------2">Continue reading on Towards AI »</a></p></div>]]></description>
            <link>https://pub.towardsai.net/remember-that-free-google-cli-that-killed-cursor-it-just-got-10x-better-e5c878d4e14e?source=rss-e432d5e4ef26------2</link>
            <guid isPermaLink="false">https://medium.com/p/e5c878d4e14e</guid>
            <category><![CDATA[google]]></category>
            <category><![CDATA[programming]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[data-science]]></category>
            <dc:creator><![CDATA[Poojan Vig]]></dc:creator>
            <pubDate>Thu, 21 Aug 2025 10:23:44 GMT</pubDate>
            <atom:updated>2025-08-22T11:25:18.417Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[The 2025 AI Engineering Report]]></title>
            <link>https://generativeai.pub/the-2025-ai-engineering-report-bf7544e6e613?source=rss-e432d5e4ef26------2</link>
            <guid isPermaLink="false">https://medium.com/p/bf7544e6e613</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[research]]></category>
            <category><![CDATA[programming]]></category>
            <category><![CDATA[software-engineering]]></category>
            <dc:creator><![CDATA[Poojan Vig]]></dc:creator>
            <pubDate>Mon, 18 Aug 2025 11:26:28 GMT</pubDate>
            <atom:updated>2025-08-20T01:24:59.056Z</atom:updated>
            <content:encoded><![CDATA[<h4>Remember when everyone said AI agents would replace developers by 2025? Well, here we are, and I’ve got some news that might surprise you.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*IB0DakFHODx4L4qYgo6IcA.png" /></figure><p>Amplify Partners just dropped their 2025 AI Engineering Report, surveying over 500 engineers who are actually building with AI in production. Not the Twitter theorists or LinkedIn thought leaders but the folks in the trenches, dealing with rate limits at 3 AM and explaining to their CEOs why the AI agent just tried to order 10,000 pizzas to the office.</p><p>Here’s the kicker: While 80% of these engineers say LLMs are working great in their day-to-day work, only 20% can say the same about AI agents. Yes, you read that right. The same agents we’ve been hyping for the past year are failing 4 out of 5 times in real production environments.</p><p>But that’s just the tip of the iceberg.</p><p>This report reveals some fascinating contradictions in our industry. Like how 70% of engineers are using RAG (Retrieval Augmented Generation) successfully, yet 31% don’t even have a proper tool to manage their prompts which they’re updating monthly, weekly, or in some cases, <em>daily</em>. Or how OpenAI has claimed 3 of the top 5 spots for customer-facing applications, while everyone’s simultaneously preparing for an open-source takeover.</p><p>What I found most intriguing? 37% of engineers who aren’t using audio AI are planning to adopt it soon. We might be standing at the edge of the voice AI revolution and not even realize it.</p><p>In this breakdown, I’ll walk you through the most important findings from this massive survey from who’s actually building AI (spoiler: even 10-year veterans are basically beginners), to why agents are struggling, to some wild predictions about the future (including that 26% of Gen Z will apparently have AI boyfriends or girlfriends by 2030).</p><p>Let’s dive into what’s really happening in AI engineering in 2025.</p><h3>Who’s Building AI in 2025?</h3><p>Here’s a wild stat that perfectly captures where we are: 45% of developers with 10+ years of experience have been working with AI for three years or less. One in ten? Less than a year.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4jI7ypIJuUyhOL3G_BEz7Q.avif" /></figure><p>Think about that. Nearly half of our industry’s most seasoned veterans are basically juniors when it comes to AI. That senior dev who’s been coding since 2010? They’re probably just as confused about vector databases as you are.</p><p>The survey pulled responses from software engineers, AI engineers, and ML specialists – though honestly, the titles are all over the place. Every company seems to have invented their own terminology for “person who argues with Claude about JSON formatting.”</p><p>But here’s what’s refreshing: we’re all figuring this out together. The field is moving so fast that even the “experts” are constantly learning. That imposter syndrome you’re feeling? Universal. The democratization is real – you don’t need a PhD anymore, just curiosity and a high tolerance for change.</p><p>Welcome to AI engineering, where everyone’s making it up as they go, and somehow, it’s working.</p><h3>The Model Landscape – OpenAI’s Dominance</h3><p>Let’s talk about the elephant in the room. When it comes to customer-facing applications, OpenAI isn’t just winning – they’re dominating.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*VahkD8hl-UXQ09jOCi0NHg.avif" /></figure><p>Three of the top five models in production are from OpenAI. Half of the top ten? Also OpenAI. While we’re all debating Claude vs GPT on Twitter, the production data tells a clear story: when real money is on the line, most companies bet on OpenAI.</p><p>But here’s where it gets interesting – 94% of teams are using LLMs for at least two different use cases, and 82% for three or more. Nobody’s building just a chatbot anymore. The same model answering customer queries in the morning is helping debug code by afternoon.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ZdNFP0AIaHPy3KmAwCwTTQ.avif" /></figure><p>The top use cases? Code generation and code intelligence lead the pack (we’re literally using AI to build more AI), followed by writing tasks. Basically, we’ve turned LLMs into the ultimate junior developer who never sleeps, never complains, and occasionally hallucinates a completely fictional Python library.</p><p>And here’s the kicker: over 50% of teams update their models monthly. Some 17% are updating weekly. We’re not just using these models – we’re in a constant dance of switching, testing, and optimizing.</p><p>The message is clear: OpenAI for production, but keep your options open. Things change fast around here.</p><h3>How Engineers Customize AI – The RAG Revolution</h3><p>If you want to know what’s actually working in production, here’s your answer: RAG. A massive 70% of engineers are using Retrieval Augmented Generation to make their AI applications actually useful.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*dLM6Mf8lSk1UCNQGx6RYMg.avif" /></figure><p>But the real surprise? 41% are fine-tuning their models. That’s way higher than I expected. Turns out, when generic models aren’t cutting it, teams aren’t afraid to get their hands dirty with LoRA, QLoRA, and good old supervised fine-tuning. The days of “just use the API” are officially over.</p><p>Now, here’s where things get chaotic. Teams are updating their prompts constantly – 70% do it monthly, and 10% are literally tweaking them daily. Daily! Imagine coming to work every morning knowing your first task is fixing whatever weird response the AI gave yesterday.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*9yTIqvfkUfoB0pU1Rl0QgQ.avif" /></figure><p>And the punchline? Despite all this constant prompt engineering, 31% of teams have absolutely no structured tool for managing these prompts. They’re probably storing them in random Notion docs, Slack messages, or that one developer’s local machine (you know who you are).</p><p>We’re essentially running production systems where the core logic changes monthly, the prompts change daily, and a third of us are managing it all with the engineering equivalent of sticky notes.</p><p>It’s both terrifying and oddly impressive that any of this works at all.</p><h3>The Multimodal Gap</h3><p>Text is eating the world. Image, audio, and video? They’re still sitting at the kids’ table.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*g_22CElS2KwHUjLkAtFJ3w.avif" /></figure><p>The production gap is stark. While everyone’s building text-based applications, the other modalities are lagging by huge margins. But here’s the plot twist: this is about to change, and audio is leading the charge.</p><p>Look at who’s planning to adopt what. Among engineers not currently using audio, 37% say they’ll implement it soon. That’s not “maybe someday” – that’s “it’s on our roadmap.” Compare that to video at just 20%, and you can see where the industry is heading.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*O-l2lw15S87fqeXpS5ocow.avif" /></figure><p>Why audio? Think about it. Voice interfaces are finally good enough. Real-time conversation APIs are here. And let’s be honest typing prompts all day is getting old. We’re about to enter the era where you’ll debug code by arguing with your AI verbally. Your open office is about to get a lot more interesting.</p><p>The companies ignoring multimodal right now are going to have a rude awakening. While they’re still perfecting their text chatbots, their competitors will be shipping voice assistants that actually work.</p><p>2025 might be remembered as the last year text dominated AI. By 2026? We’ll all be talking to our computers, and wondering why it took us so long.</p><h3>The Agent Reality Check</h3><p>Remember all those “AI agents will revolutionize everything” predictions from last year? Time for a reality check.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6tcjCLnQ8mus44IAK9CCtw.avif" /></figure><p>While 80% of engineers say LLMs are working great, only 20% can say the same about agents. That’s not a typo. Agents are failing four out of five times in production. The autonomous AI revolution? Still buffering.</p><p>But here’s the thing nobody’s giving up. Less than 10% of teams say they have no plans for agents. Everyone else is either building them, fixing them, or planning to deploy them. We’re all convinced agents are the future; we just can’t get them to work reliably in the present.</p><p>The trust issue becomes crystal clear when you look at permissions. Most production agents have write access, but with heavy human oversight. Only 13% are allowed to act completely independently. We’re basically treating agents like intern developers sure, you can make changes, but someone’s definitely reviewing your PR.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*GETDaGU5J_rVc1_bn43lKw.avif" /></figure><p>What’s holding agents back? They’re great at demos where everything’s controlled. But production? That’s where they meet edge cases, ambiguous instructions, and users who definitely didn’t read the documentation. Turns out, giving an AI system decision-making power is harder than making it sound smart in a chatbot.</p><p>The gap between agent hype and agent reality might be the biggest story in AI right now. We’re not abandoning them we’re just learning that the path from “cool demo” to “production system” is longer than anyone expected.</p><h3>Engineering Challenges &amp; Infrastructure</h3><p>Let’s talk about the unglamorous stuff that actually makes AI work in production.</p><p>First up: vector databases. 65% of teams are using dedicated vector databases rather than trying to make PostgreSQL do tricks it wasn’t meant to do. Turns out, when you’re dealing with embeddings at scale, specialized tools actually matter. Who would’ve thought?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*gVQy6K96kiukvEwtUQ8zwQ.png" /></figure><p>LangChain is absolutely dominating the framework game — it’s twice as popular as the next option. Love it or hate it (and trust me, developers have opinions), it’s become the default way to build LLM applications. It’s like the React of AI development: complicated, sometimes overkill, but everyone knows it.</p><p>But here’s the reality check on evaluation: despite all our fancy tools, human review is still king. Most teams use a combination of benchmarks, user feedback, and good old-fashioned “does this look right?” to validate their systems. We’re basically QA-testing intelligence itself, and nobody’s quite figured out how to automate that yet.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*MbdZL1LltLVnC3UHThHINw.avif" /></figure><p>The top pain points? Accuracy (obviously), latency (because users won’t wait), and cost (those API bills add up fast). One engineer summed it up perfectly: “We’re trying to build reliable systems on top of probabilistic models that charge by the token.”</p><p>The infrastructure is maturing, but we’re still duct-taping solutions together. Every team has their own Frankenstein stack of monitoring tools, evaluation frameworks, and that one crucial Python script nobody wants to touch.</p><p>Welcome to AI engineering: where cutting-edge technology meets “whatever works.”</p><h3>Surprising Predictions &amp; Industry Beliefs</h3><p>Alright, let’s get into the spicy stuff what AI engineers really think about the future.</p><p>First, the prediction that made me do a double-take: engineers believe 26% of Gen Z will have AI boyfriends or girlfriends by 2030. That’s more than one in four. We went from “AI can help you write emails” to “AI might be your next relationship” real quick.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*sPQm3ekfbB29ctj6fje6Hw.avif" /></figure><p>On the technical side, there’s surprising optimism about open source. Most respondents believe open-source models will eventually match closed-source performance. The OpenAI moat? Maybe not as deep as we thought. This isn’t just wishful thinking teams are actively preparing for a world where they can run GPT-4 level models on their own infrastructure.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*UQoEIr7w_Oc7VlR_Lc2auQ.avif" /></figure><p>Here’s another hot take: over 40% of engineers won’t pay extra for reasoning models or longer inference times. Despite all the hype around o1 and similar models, teams are saying “regular LLMs are fine, thanks.” Speed and cost still beat marginal intelligence gains.</p><p>And my favorite finding? Many engineers disagree with “attention is all you need.” After years of transformer dominance, there’s growing skepticism that we’ve found the final architecture. People are ready for the next breakthrough.</p><p>The community is beautifully divided on almost everything. We can’t agree if agents will work, if open source will win, or if our future romantic partners will be powered by neural networks.</p><p>The only consensus? Nobody really knows what’s coming next, and that’s what makes this field so exciting.</p><h3>Key Takeaways for Developers</h3><p>So what does all this data mean for you? Here’s your practical playbook based on what’s actually working in production.</p><p><strong>Start with RAG, not agents.</strong> 70% success rate vs 20%? The math is clear. Get RAG working first, then worry about autonomous systems. Walk before you run, even if the running demos look cooler.</p><p><strong>Prepare for constant change.</strong> If you’re not updating models monthly and prompts weekly, you’re already behind. Build your systems to handle change, not resist it. That means version control for prompts (yes, really) and abstraction layers for model swapping.</p><p><strong>Audio is coming =&gt; be ready.</strong> With 37% planning adoption, voice interfaces are about to explode. Start experimenting now, even if it’s just a side project. The teams that figure out voice UX early will have a massive advantage.</p><p><strong>Don’t skip the boring stuff.</strong> Get a proper vector database. Set up real monitoring. Build evaluation pipelines. The infrastructure isn’t sexy, but it’s what separates demos from products.</p><p><strong>Keep your options open.</strong> Yes, OpenAI dominates today, but things change fast. Build with abstraction layers that let you switch models without rewriting everything. The best model in December might not be the best in January.</p><p><strong>Human-in-the-loop is not a failure.</strong> It’s the current best practice. Don’t feel bad about having humans review AI outputs literally everyone is doing it. Pure automation is a goal, not a requirement.</p><p>Most importantly: <strong>everyone is figuring this out together.</strong> That senior engineer with 15 years experience? They’re just as confused by the latest paper as you are. Stay curious, ship fast, and don’t let imposter syndrome stop you from building.</p><h3>Conclusion: The State of AI Engineering</h3><p>We’re in the messy middle of the AI revolution. LLMs work great, agents mostly don’t. Everyone’s using RAG but managing prompts with sticky notes. OpenAI dominates while we prepare for open source.</p><p>The most striking finding from Amplify Partners’ 2025 AI Engineering Report? How universal these struggles are. Whether you’re at a startup or Big Tech, we’re all fighting the same battles: agents that won’t cooperate, daily prompt updates, and infrastructure held together with Python scripts and prayer.</p><p>But nobody’s giving up. Despite agents failing 80% of the time, we keep building. Despite the chaos, we keep shipping. We know this technology is transformative we just need to make it boring and reliable.</p><p>2025 isn’t the year AI takes over. It’s the year we learned what AI can actually do versus what we hoped it would do. And that’s exactly the reality check we needed.</p><p><em>What surprised you most? Let me know in the comments.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/700/0*ZzfY8_3v2IAqcnti.png" /></figure><p>This story is published on <a href="https://generativeai.pub/">Generative AI</a>. Connect with us on <a href="https://www.linkedin.com/company/generative-ai-publication">LinkedIn</a> and follow <a href="https://www.zeniteq.com/">Zeniteq</a> to stay in the loop with the latest AI stories.</p><p>Subscribe to our <a href="https://www.generativeaipub.com/">newsletter</a> and <a href="https://www.youtube.com/@generativeaipub">YouTube</a> channel to stay updated with the latest news and updates on generative AI. Let’s shape the future of AI together!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/700/0*XQC8c0rzWDA_Wl2B.png" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=bf7544e6e613" width="1" height="1" alt=""><hr><p><a href="https://generativeai.pub/the-2025-ai-engineering-report-bf7544e6e613">The 2025 AI Engineering Report</a> was originally published in <a href="https://generativeai.pub/">Generative AI</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Same AI, Different Prompts: The 500% Performance Gap Nobody Talks About]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://pub.towardsai.net/same-ai-different-prompts-the-500-performance-gap-nobody-talks-about-a896f405354b?source=rss-e432d5e4ef26------2"><img src="https://cdn-images-1.medium.com/max/2600/1*KDNHPaTLt6QoGPZtTcp7Lg.jpeg" width="2944"></a></p><p class="medium-feed-snippet">Two developers. Same AI model. Same task: &#x201C;Build a user authentication system.&#x201D;</p><p class="medium-feed-link"><a href="https://pub.towardsai.net/same-ai-different-prompts-the-500-performance-gap-nobody-talks-about-a896f405354b?source=rss-e432d5e4ef26------2">Continue reading on Towards AI »</a></p></div>]]></description>
            <link>https://pub.towardsai.net/same-ai-different-prompts-the-500-performance-gap-nobody-talks-about-a896f405354b?source=rss-e432d5e4ef26------2</link>
            <guid isPermaLink="false">https://medium.com/p/a896f405354b</guid>
            <category><![CDATA[programming]]></category>
            <category><![CDATA[case-study]]></category>
            <category><![CDATA[prompt]]></category>
            <category><![CDATA[prompt-engineering]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Poojan Vig]]></dc:creator>
            <pubDate>Tue, 12 Aug 2025 13:02:09 GMT</pubDate>
            <atom:updated>2025-08-12T13:02:09.963Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Claude Code: Best Practices]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://generativeai.pub/claude-code-best-practices-076383bc2569?source=rss-e432d5e4ef26------2"><img src="https://cdn-images-1.medium.com/max/1612/1*FWSfUeBBhrE2ni2E-GevJw.png" width="1612"></a></p><p class="medium-feed-link"><a href="https://generativeai.pub/claude-code-best-practices-076383bc2569?source=rss-e432d5e4ef26------2">Continue reading on Generative AI »</a></p></div>]]></description>
            <link>https://generativeai.pub/claude-code-best-practices-076383bc2569?source=rss-e432d5e4ef26------2</link>
            <guid isPermaLink="false">https://medium.com/p/076383bc2569</guid>
            <category><![CDATA[programming]]></category>
            <category><![CDATA[claude]]></category>
            <category><![CDATA[open-source]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[generative-ai-tools]]></category>
            <dc:creator><![CDATA[Poojan Vig]]></dc:creator>
            <pubDate>Mon, 04 Aug 2025 04:25:07 GMT</pubDate>
            <atom:updated>2025-08-04T07:55:30.462Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[JSON Prompting for Every AI: One Format to Rule GPT, Claude & Gemini]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://generativeai.pub/json-prompting-for-every-ai-one-format-to-rule-gpt-claude-gemini-6f276127e112?source=rss-e432d5e4ef26------2"><img src="https://cdn-images-1.medium.com/max/1920/1*4xChjAReU51Yl0V96purGA.jpeg" width="1920"></a></p><p class="medium-feed-snippet">You asked ChatGPT for a &#x201C;short social media post&#x201D; and got a 300-word essay. You prompted Claude for &#x201C;5 key insights&#x201D; and received 8&#x2026;</p><p class="medium-feed-link"><a href="https://generativeai.pub/json-prompting-for-every-ai-one-format-to-rule-gpt-claude-gemini-6f276127e112?source=rss-e432d5e4ef26------2">Continue reading on Generative AI »</a></p></div>]]></description>
            <link>https://generativeai.pub/json-prompting-for-every-ai-one-format-to-rule-gpt-claude-gemini-6f276127e112?source=rss-e432d5e4ef26------2</link>
            <guid isPermaLink="false">https://medium.com/p/6f276127e112</guid>
            <category><![CDATA[claude]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[gpt]]></category>
            <category><![CDATA[prompt-engineering]]></category>
            <category><![CDATA[gemini]]></category>
            <dc:creator><![CDATA[Poojan Vig]]></dc:creator>
            <pubDate>Sat, 02 Aug 2025 11:13:32 GMT</pubDate>
            <atom:updated>2025-08-02T13:41:45.896Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Qwen Code CLI + Qwen3-Coder Let’s Set Up Qwen Code, Better than Claude Code?]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://generativeai.pub/qwen-code-cli-qwen3-coder-lets-set-up-qwen-code-better-than-claude-code-3ada7b00dd1c?source=rss-e432d5e4ef26------2"><img src="https://cdn-images-1.medium.com/max/2600/1*Ge3THkPFhhLV51SeyjlHAQ.jpeg" width="3202"></a></p><p class="medium-feed-snippet">24 hours ago, I was paying $20/month for Claude Code without question. Then Alibaba dropped a bombshell.</p><p class="medium-feed-link"><a href="https://generativeai.pub/qwen-code-cli-qwen3-coder-lets-set-up-qwen-code-better-than-claude-code-3ada7b00dd1c?source=rss-e432d5e4ef26------2">Continue reading on Generative AI »</a></p></div>]]></description>
            <link>https://generativeai.pub/qwen-code-cli-qwen3-coder-lets-set-up-qwen-code-better-than-claude-code-3ada7b00dd1c?source=rss-e432d5e4ef26------2</link>
            <guid isPermaLink="false">https://medium.com/p/3ada7b00dd1c</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[generative-ai-tools]]></category>
            <category><![CDATA[claude]]></category>
            <category><![CDATA[coding]]></category>
            <category><![CDATA[openai]]></category>
            <dc:creator><![CDATA[Poojan Vig]]></dc:creator>
            <pubDate>Fri, 25 Jul 2025 05:25:09 GMT</pubDate>
            <atom:updated>2025-07-25T05:25:09.279Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Google Just Killed $200/Month AI Coding Tools With This Free Terminal Assistant]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://pub.towardsai.net/google-just-killed-200-month-ai-coding-tools-with-this-free-terminal-assistant-cdb4051cacf8?source=rss-e432d5e4ef26------2"><img src="https://cdn-images-1.medium.com/max/1920/1*C8IIqiREYvCTevpArUBf4Q.png" width="1920"></a></p><p class="medium-feed-snippet">On June 25, 2025, Google dropped a nuclear bomb on the AI coding assistant market.</p><p class="medium-feed-link"><a href="https://pub.towardsai.net/google-just-killed-200-month-ai-coding-tools-with-this-free-terminal-assistant-cdb4051cacf8?source=rss-e432d5e4ef26------2">Continue reading on Towards AI »</a></p></div>]]></description>
            <link>https://pub.towardsai.net/google-just-killed-200-month-ai-coding-tools-with-this-free-terminal-assistant-cdb4051cacf8?source=rss-e432d5e4ef26------2</link>
            <guid isPermaLink="false">https://medium.com/p/cdb4051cacf8</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[vibe-coding]]></category>
            <category><![CDATA[google]]></category>
            <category><![CDATA[claude]]></category>
            <category><![CDATA[gemin]]></category>
            <dc:creator><![CDATA[Poojan Vig]]></dc:creator>
            <pubDate>Wed, 02 Jul 2025 00:02:52 GMT</pubDate>
            <atom:updated>2025-07-02T00:02:52.543Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[The Illusion of Thinking: Apple’s Groundbreaking Study on Why Reasoning AI Models Break Down]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://generativeai.pub/the-illusion-of-thinking-apples-groundbreaking-study-on-why-reasoning-ai-models-break-down-a7862a98da5d?source=rss-e432d5e4ef26------2"><img src="https://cdn-images-1.medium.com/max/1536/1*uDYEBYBlx48WYI-1eRZusg.png" width="1536"></a></p><p class="medium-feed-snippet">So Apple just published this fascinating paper about how AI &#x201C;thinking&#x201D; models work (or don&#x2019;t work, actually)</p><p class="medium-feed-link"><a href="https://generativeai.pub/the-illusion-of-thinking-apples-groundbreaking-study-on-why-reasoning-ai-models-break-down-a7862a98da5d?source=rss-e432d5e4ef26------2">Continue reading on Generative AI »</a></p></div>]]></description>
            <link>https://generativeai.pub/the-illusion-of-thinking-apples-groundbreaking-study-on-why-reasoning-ai-models-break-down-a7862a98da5d?source=rss-e432d5e4ef26------2</link>
            <guid isPermaLink="false">https://medium.com/p/a7862a98da5d</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[thinking]]></category>
            <category><![CDATA[apple]]></category>
            <category><![CDATA[llm]]></category>
            <dc:creator><![CDATA[Poojan Vig]]></dc:creator>
            <pubDate>Tue, 17 Jun 2025 06:12:56 GMT</pubDate>
            <atom:updated>2025-06-17T06:12:56.158Z</atom:updated>
        </item>
    </channel>
</rss>