<?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 Prathamesh Khade on Medium]]></title>
        <description><![CDATA[Stories by Prathamesh Khade on Medium]]></description>
        <link>https://medium.com/@prathamesh.khade20?source=rss-9af5193a2117------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*-LpbXsqUwsj-poxCLbfXBQ.jpeg</url>
            <title>Stories by Prathamesh Khade on Medium</title>
            <link>https://medium.com/@prathamesh.khade20?source=rss-9af5193a2117------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 08:41:43 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@prathamesh.khade20/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[The Personal AI Era Has Arrived — And It Isn’t the Smartest Model in the Room]]></title>
            <link>https://medium.com/@prathamesh.khade20/the-personal-ai-era-has-arrived-and-it-isnt-the-smartest-model-in-the-room-7b4f72f371d0?source=rss-9af5193a2117------2</link>
            <guid isPermaLink="false">https://medium.com/p/7b4f72f371d0</guid>
            <category><![CDATA[llm-finetuning]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[inklings]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[thinking-machines-lab]]></category>
            <dc:creator><![CDATA[Prathamesh Khade]]></dc:creator>
            <pubDate>Sat, 18 Jul 2026 18:50:03 GMT</pubDate>
            <atom:updated>2026-07-18T18:50:03.958Z</atom:updated>
            <content:encoded><![CDATA[<h3>Why Thinking Machines Lab’s first model, Inkling, matters more for what it admits than for what it wins</h3><p>Every few months, a new “state of the art” model launches with a chart showing it beating everyone else. Inkling, the debut model from Mira Murati’s Thinking Machines Lab, does the opposite. Buried in its own launch materials is a sentence you almost never see from an AI lab: <em>this is not the strongest model available today, open or closed.</em></p><p>That’s not false modesty. It’s a strategy. And understanding why a $12B startup, founded by OpenAI’s former CTO, would ship something that isn’t #1 tells you more about where AI is heading in 2026 than any leaderboard does.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*JX7XjOfS_b07RwcJPqTaBg.png" /></figure><h3>1. Who’s Behind This, and Why It’s a Big Deal</h3><p>Thinking Machines Lab was founded in February 2025 by Mira Murati — the person who ran the technology behind ChatGPT and briefly served as OpenAI’s interim CEO during the chaotic November 2023 board upheaval. She brought a group of senior OpenAI researchers with her, including co-founder John Schulman and former OpenAI VP Lilian Weng.</p><p>Before shipping a single product, the company raised a <strong>record $2 billion seed round at a $12 billion valuation</strong>, backed by Nvidia, AMD, Cisco, a16z, and Jane Street. For over a year, the only public product was <strong>Tinker</strong>, a fine-tuning API launched in October 2025. Skeptics wondered whether the hype would ever turn into a model.</p><p>In July 2026, it did. <strong>Inkling</strong> is Thinking Machines’ first model trained completely from scratch — and the story around it is far more interesting than its benchmark scores.</p><h3>2. Meet Inkling: The Specs That Matter</h3><p>Inkling is a <strong>Mixture-of-Experts (MoE) transformer</strong> — one of the largest open models in existence:</p><ul><li><strong>975 billion total parameters</strong>, but only <strong>~41 billion active</strong> per prompt (MoE routing keeps it fast and cheap to run)</li><li><strong>1 million token context window</strong></li><li><strong>Pretrained on 45 trillion tokens</strong> spanning text, images, audio, and video</li><li><strong>Natively multimodal</strong> — it transcribes speech, follows spoken instructions, reads charts and diagrams, and can call a Python tool mid-thought to zoom and crop into an image while reasoning</li><li><strong>Outputs are text-only</strong>, but that covers code, structured data, and styled artifacts</li><li>Released <strong>fully open-weight under Apache 2.0</strong> — free to download, run, and fine-tune, with weights live on Hugging Face in both a standard checkpoint and an NVFP4 version optimized for Nvidia Blackwell hardware</li><li>Trained on Nvidia’s GB300 NVL72 systems, going from <strong>zero to release in under nine months</strong> — a fraction of the multi-year cycles typical at labs like OpenAI and Anthropic</li></ul><p>Alongside the flagship, Thinking Machines previewed <strong>Inkling-Small</strong>, a 276B-parameter, 12B-active model that already matches or beats its bigger sibling on several benchmarks — a sign that the “small, personal model” trend is only accelerating.</p><p><em>(Under the hood, Inkling’s MoE design — 256 routed experts plus 2 shared experts per layer, sigmoid routing, sliding-window/global attention at a 5:1 ratio — closely follows the architecture DeepSeek popularized with V3. More on why that matters below.)</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*J3ewvZmvp2tJQ-hZjqX9Nw.gif" /></figure><h3>3. The Honest Benchmarks</h3><p>Thinking Machines doesn’t hide where Inkling falls short of the frontier:</p><p>Benchmark Inkling Best Comparable Model Humanity’s Last Exam (text-only) 29.7% GLM 5.2–40.1% SWE-bench Verified 77.6% Behind current leaders Terminal-Bench 2.1 63.8% GLM 5.2–82.7% Artificial Analysis Intelligence Index 41 —</p><p>That last number is the real headline: Inkling debuts as the <strong>leading open-weight model built by a U.S. lab</strong>, three points ahead of Nvidia’s Nemotron 3 Ultra and comfortably ahead of Meta’s Gemma 4 and OpenAI’s gpt-oss-120b. It’s not the world’s smartest model — it’s the strongest <em>American, freely downloadable</em> one.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*3Ivj19GX44RBYwgC0qM5wQ.png" /></figure><p>Where it does shine is <strong>efficiency and breadth</strong>. On Terminal-Bench, Inkling matches Nemotron 3 Ultra’s coding performance while spending roughly a third of the tokens. It also holds its own against much larger closed models on agentic web-development tasks judged by blinded human raters — landing close to Claude Opus-class performance and ahead of Kimi K2.6.</p><p>The design philosophy is explicit: instead of chasing one leaderboard, Inkling was trained as a <strong>balanced generalist</strong> — agentic tasks, reasoning, coding, instruction-following, factuality, vision, and audio — with a <strong>controllable “thinking effort” dial</strong> (0.2 to 0.99) that lets developers trade speed for accuracy on demand.</p><h3>4. The Feature Nobody Else Is Betting On: Knowing What It Doesn’t Know</h3><p>This is the part worth slowing down for.</p><p>Most models hallucinate confidently. Thinking Machines trained Inkling with reinforcement learning against <strong>proper scoring rules</strong> over a large corpus of resolved real-world questions — essentially teaching the model to be <em>calibrated</em>, not just fluent.</p><p>The result: on ForecastBench (without search), Inkling posts a Brier score of 61.1 — beating GPT-5.5 (59.1) and Claude Opus 4.8 (54.6), and tying Gemini 3.1 Pro. In practice, this means that when you ask Inkling something the evidence doesn’t clearly support, it hedges and attaches a probability instead of inventing an answer.</p><p><strong>Example from testing:</strong> given a single prompt — <em>“research the current state of open-weight models, make three concrete predictions about where the field lands by year-end, and build an interactive forecast dashboard”</em> — Inkling searched the web, pulled real numbers, explicitly flagged where evidence was thin, assigned confidence levels to each prediction, and then built a working dashboard (cards, animated confidence bars, sorting, dark UI) — all in one response, no follow-up prompt needed.</p><p>This kind of <strong>epistemic humility</strong> was reinforced with two automated graders during training: a rubric grader checking answers against what a good response should contain, and a claims grader that runs agentic web search to fact-check every claim and penalize the ones that don’t hold up.</p><p>Safety got similar rigor: internal evals covered CBRN, cyber, and loss-of-control risks alongside human-facing risks like sycophancy and manipulation, verified by external testers. On the Fortress benchmark (weapons/violence refusal vs. benign lookalikes), Inkling scored 78% on adversarial prompts — the strongest of any open-weight model tested — while still answering 95.9% of harmless lookalike questions correctly.</p><h3>5. The Real Innovation Isn’t the Model — It’s the Business Model</h3><p>Here’s the strategic bet: <strong>Thinking Machines isn’t selling access to intelligence. It’s selling a base to build on.</strong></p><p>There’s no metered, per-token API play here. Revenue comes from <strong>Tinker</strong>, the fine-tuning platform, with Inkling positioned as raw material you customize and run on your own infrastructure.</p><p>The proof point is <strong>Bridgewater Associates</strong>, the hedge fund, which used Tinker to fine-tune a lightweight version of Alibaba’s Qwen on proprietary financial data. The result scored 84.7% on leading financial-reasoning benchmarks — beating top proprietary models at under 10% of the cost.</p><p>That’s the pitch in one sentence: <strong>a small model fine-tuned on your data can beat a giant generalist in your specific niche, for pennies.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/678/1*uCV2YYWPF9gAZE1ai4sl-w.jpeg" /></figure><p>This is why so many organizations were already drifting toward Chinese open models — cheap, excellent, and fine-tunable. At one point roughly <strong>45% of enterprise tokens</strong> routed through OpenRouter went to Chinese models. Coinbase reportedly cut its AI bill nearly in half by moving agents to GLM and Kimi. Cursor built its Composer model on Kimi.</p><p>But Washington is closing that door — the State Department has warned companies about Chinese models, congressional committees have opened probes, and procurement bans are being drafted. Meanwhile Meta quietly deprioritized Llama after Llama 4 underperformed, leaving a gap in the Western open-source ecosystem. <strong>Inkling exists to fill that vacuum</strong> — not the best base model on Earth, but the best <em>permitted</em> one for anyone operating under U.S. policy constraints.</p><p>Analysts are already framing this as a structural decision, not a technical one: choosing a base model today is an architecture decision, because fine-tuning costs compound with every downstream customization.</p><h3>6. The Uncomfortable Irony</h3><p>Here’s the twist that’s harder to spin: Inkling’s own architecture largely follows <strong>DeepSeek </strong>V3’s MoE blueprint, and part of its training data comes from <strong>synthetic outputs of Kimi K2.5</strong>, a Chinese open model. An American lab, staffed by people who built ChatGPT, chose to build its flagship on Chinese open-source foundations rather than anything resembling its founders’ former employer’s approach.</p><p>For the past year, U.S. officials and OpenAI itself have accused Chinese labs of “distilling” American frontier models without permission — treating it as intellectual theft. Now the flow runs the other direction, and it’s simply called engineering. To be fair, the Chinese models involved here are openly and permissively licensed, so nothing improper happened. But the rhetorical double standard is hard to miss — and it’s a live signal about how thin the moat between “open” ecosystems has become, regardless of geography.</p><h3>7. What This Actually Means, Depending on Who You Are</h3><p><strong>If you’re a student or early-career engineer:</strong> this is the best time in years to build something real without a massive compute budget. Download the weights, fine-tune a specialist model on Tinker, and you can outperform frontier generalists in a narrow domain — a thesis project, a hackathon tool, a niche SaaS idea — for a fraction of the cost of calling GPT or Claude at scale.</p><p><strong>If you’re an AI practitioner or researcher:</strong> the epistemic-calibration training recipe (proper scoring rules + claims-verification grading) is arguably more reusable than the model itself. Expect this pattern to show up across labs within a year.</p><p><strong>If you’re in product or leadership:</strong> the message is blunt — picking a foundation model is no longer just a cost decision, it’s a strategic and increasingly geopolitical one. Fine-tuned models inherit the ceiling of their base. Teams building on the strongest available open foundation will out-execute teams walled off onto weaker “permitted” alternatives — regardless of talent or budget.</p><h3>8. The Bigger Picture</h3><p>Inkling probably won’t top a single leaderboard for long. That was never the point. It’s a signal that the next phase of AI isn’t about one company’s model being smartest — it’s about <strong>which base you build your organization’s intelligence on, and who owns the result.</strong></p><p>The frontier lab race made headlines for three years. The next three years might be about something quieter but more consequential: the shift from renting intelligence by the token to <strong>owning a model you shaped yourself.</strong> That’s the personal AI era. Inkling is simply the first serious American entry into it.</p><p><em>Sources: Thinking Machines Lab launch materials, Artificial Analysis, TechCrunch, MarkTechPost, Databricks, TechTimes — July 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7b4f72f371d0" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Multi agent system for content creation]]></title>
            <link>https://medium.com/@prathamesh.khade20/multi-agent-system-for-content-creation-aaefa5350012?source=rss-9af5193a2117------2</link>
            <guid isPermaLink="false">https://medium.com/p/aaefa5350012</guid>
            <category><![CDATA[ai-agent]]></category>
            <dc:creator><![CDATA[Prathamesh Khade]]></dc:creator>
            <pubDate>Sun, 14 Jul 2024 20:44:45 GMT</pubDate>
            <atom:updated>2024-07-15T04:19:36.933Z</atom:updated>
            <content:encoded><![CDATA[<h3>Multi-Agent Systems in AI: A Comprehensive Overview</h3><p>In the evolving field of artificial intelligence (AI) and language models, the concept of a multi-agent system stands out. A multi-agent system involves several autonomous agents, each driven by language models, collaborating to achieve a common goal. Each agent is endowed with a unique persona, role, and context defined by specific prompts within the language model. These agents have access to various tools to execute their tasks. By working together, these agents bring diverse perspectives, leading to enhanced decision-making and more robust outcomes.</p><h3>Differences Between Multi-Agent and Single-Agent Systems</h3><figure><img alt="Single agent system vs Multi agent system" src="https://cdn-images-1.medium.com/max/300/1*hwXgEstrQ9nwixGQK9JckA.jpeg" /><figcaption>Single agent system vs Multi agent system</figcaption></figure><p>The fundamental difference between multi-agent and single-agent systems lies in their approach to decision-making and interaction. In single-agent systems, a central agent makes all decisions, while other agents act as subordinate entities. This centralized approach might miss out on multiple perspectives and possibilities. In contrast, multi-agent systems consist of multiple intelligent agents, each capable of making independent decisions and influencing the environment. This decentralization allows for more dynamic and adaptive responses to complex problems.</p><h3>The Concept of Multi-Agent Architecture</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/673/1*jDrfixstw6-KY8ngytccKQ.png" /><figcaption>Multi agent system</figcaption></figure><p>The core idea behind multi-agent architecture is to create agents with distinct contexts, enabling them to bring different perspectives based on their roles. Although these agents might utilize the same large language model (LLM), their specific roles, goals, and contexts ensure diverse behaviors, akin to team members in a collaborative environment. For example, one agent might generate application code, while another reviews it. Through iterative interactions, these agents can refine the code, producing superior results and minimizing issues like hallucinations and bias. Effective prompt engineering is crucial in defining impactful multi-agent applications.</p><h3>Advantages of Multi-Agent system</h3><h4>1) Separation of Concerns</h4><p>One of the primary benefits of multi-agent designs is the separation of concerns. Each agent operates with specific instructions and few-shot examples, leveraging separate fine-tuned language models and tools. This division of labor leads to better outcomes, as each agent focuses on a specialized task rather than juggling multiple responsibilities. By isolating tasks, agents can perform more efficiently and effectively.</p><h4>2) Modularity</h4><p>Multi-agent designs offer a modular approach to problem-solving. By breaking down complex problems into manageable units, specialized agents and language models can tackle specific tasks. This modularity allows for independent evaluation and improvement of each agent without disrupting the entire system.</p><h4>3) Diversity</h4><p>Diversity within agent teams is another significant advantage. By incorporating varied perspectives, multi-agent systems can enhance the quality of the output and reduce issues such as hallucinations and bias. This diversity mirrors a typical human team, where different viewpoints contribute to more comprehensive and balanced decisions.</p><h4>4) Reusability</h4><p>Once created, agents can be reused across various use cases, contributing to an ecosystem of agents capable of solving problems collaboratively through a proper orchestration framework (such as AutoGen, Crew.ai, etc.). This reusability enhances the efficiency and scalability of multi-agent systems, allowing developers to leverage existing agents for new applications without starting from scratch.</p><h4>5) Enhanced Problem Solving</h4><p>Multi-agent systems can tackle complex problems more effectively than single-agent systems. By leveraging the diverse perspectives and specialized skills of multiple agents, these systems can explore a broader range of solutions and identify the most effective strategies.</p><h4>6) Scalability</h4><p>Multi-agent systems are inherently scalable. As the complexity or scope of a problem increases, additional agents can be introduced to handle new tasks or responsibilities. This scalability makes multi-agent systems suitable for a wide range of applications, from simple tasks to highly complex, dynamic environments.</p><h3>Essential Components of Multi-Agent Architecture</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/664/1*JXXtIXRKZI_dhSspygejbw.png" /></figure><h4>A) Agents</h4><p>Intelligent agents with clear roles and contexts operate on an LLM. Each agent is designed to perform specific tasks and contribute to the overall goal of the multi-agent system.</p><h4># Key Elements of AI Agents</h4><p>When designing AI agents for a multi-agent system, several key elements must be considered to ensure effective performance and collaboration. These elements define how the agents operate, interact, and achieve their goals.</p><h4>1. Role Playing</h4><p>Role playing involves defining a specific role for each agent. For example, an agent could be assigned the role of a financial analyst, tasked with providing an analysis of Tesla’s stock. Role playing helps agents focus on their specific tasks and responsibilities, ensuring that each agent contributes effectively to the overall system.</p><h4>2. Focus</h4><p>Agents should be focused on specific tasks to leverage multiple agents for various purposes. By concentrating on individual tasks, agents can perform more efficiently and deliver better results.</p><h4>3. Tools</h4><p>Each agent should be equipped with the necessary tools to perform its tasks. These tools can range from data retrieval systems to specialized software for analysis and content generation. Proper tools empower agents to execute their tasks effectively.</p><h4>4. Cooperation</h4><p>Cooperation among agents is crucial for a multi-agent system. This involves sharing task feedback, delegating tasks when necessary, and working together to achieve common goals. Effective cooperation ensures that agents can support each other and enhance overall system performance.</p><h4>5. Guardrails</h4><p>Implementing guardrails is essential to prevent agents from providing repeated or irrelevant answers. Guardrails help maintain the quality and relevance of the agents’ outputs by ensuring that they adhere to predefined guidelines and standards.</p><h4>6. Memory</h4><p>Memory is a critical element for AI agents, encompassing both short-term and long-term memory capabilities:</p><ul><li><strong>Short-Term Memory</strong>: This involves memory during task execution, allowing agents to share knowledge, activities, and learnings with other agents in real time.</li><li><strong>Long-Term Memory</strong>: Long-term memory is stored in a database, enabling agents to retain information over extended periods. This leads to the development of ‘self-improving’ agents that can learn from past experiences and improve their performance over time.</li><li><strong>Entity Memory</strong>: Entity memory pertains to storing information about persons, organizations, and locations. This helps agents maintain context and relevance in their interactions and outputs.</li></ul><h4>B) Connections</h4><p>The connections between agents define how they interact and share information. Effective communication channels are essential for collaboration and coordination among agents.</p><h4>C) Orchestration</h4><p>Orchestration is the process by which agents work together. This can be sequential, hierarchical, bi-directional, or any other structured interaction model. Proper orchestration ensures that agents collaborate effectively, leveraging each other’s strengths to achieve the desired outcome.</p><h4>D) Human Interaction</h4><p>Human involvement is often necessary for decision-making and result evaluation. Humans can provide oversight, make critical decisions, and intervene when necessary to guide the multi-agent system toward its goals.</p><h4>E) Tools</h4><p>Tools are essential for agents to perform specific tasks, such as searching the web for information, generating or reading documents, or uploading generated code to platforms like GitHub. Access to the right tools enhances the capabilities of agents and allows them to perform their tasks more effectively.</p><h4># Key Elements of an Effective Agent Tool</h4><p>Creating an effective AI agent tool involves ensuring it is:</p><ul><li><strong>Versatile</strong>: Capable of handling various tasks and adapting to different requirements.</li><li><strong>Fault-Tolerant</strong>: Able to manage errors and continue functioning effectively.</li><li><strong>Efficient</strong>: Utilizing caching mechanisms to prevent unnecessary requests, stay within rate limits, and save time with faster-cached results.</li></ul><h4># Task performed by tools:</h4><p>There are variety of tools that agents can use to perform their tasks like:</p><ol><li>Search the Internet</li><li>Scrape Websites</li><li>Connect to Databases</li><li>Call APIs</li><li>Send Notifications</li></ol><h4># Specific Tools</h4><ol><li>Google SERP search</li><li>Website scraper</li><li>Directory read</li><li>File read</li><li>Selenium scraper</li><li>Directory RAG search</li><li>PDF RAG search</li><li>TXT RAG search</li><li>CSV RAG search</li><li>XML RAG search</li><li>JSON RAG search</li><li>DOCX RAG search</li><li>MDX RAG search</li><li>PG RAG search</li><li>Website RAG search</li><li>GitHub RAG search</li><li>CodeDocs RAG search</li><li>YouTube video RAG search</li><li>Channel RAG search</li></ol><h4>F) LLM (Large Language Model)</h4><p>The LLM serves as the backbone of the multi-agent system, providing the foundational language model that supports agent inference. The choice of LLM can significantly impact the performance and capabilities of the agents.</p><h3>Frameworks for Building Multi-Agent Applications</h3><p>Several frameworks facilitate the development of multi-agent applications. While the field is rapidly evolving, these frameworks simplify building and managing multiple agents. Here are some notable examples:</p><h4>1) CrewAI</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/383/1*9vtKExZ07iU8jx-JCovkDQ.png" /></figure><p>CrewAI is an emerging framework that orchestrates role-playing, autonomous AI agents. CrewAI promotes collaborative intelligence, enabling agents to work together seamlessly on complex tasks. It allows AI agents to assume roles, share goals, and operate as a cohesive unit. This framework is gaining popularity and is a favorite for building multi-agent applications due to its flexibility and focus on collaborative intelligence.</p><h4>2) OpenAI Assistant</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/318/1*sr7o-eqJEHcmmIuqUcwQHQ.jpeg" /></figure><p>OpenAI Assistant is one of the pioneering frameworks supporting multi-agent architecture. It enables the creation of persistent, multimodal multi-agent systems that interact with users over time. Agents can access files, tools, and communicate with each other for task execution. This framework is ideal for applications requiring long-term collaboration and interaction.</p><h4>3) Autogen</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/310/1*5AuYpBkWrnxhQ6vTh3PJrQ.jpeg" /></figure><p>Autogen, an emerging framework by Microsoft, offers an open-source platform with an intuitive UI (Autogen Studio) for building robust multi-agent applications. It employs a well-defined orchestrator-based approach to multi-agent architecture, using LLMs for reasoning and action, augmented with custom information sources. Autogen provides a versatile and user-friendly environment for developing and managing multi-agent systems.</p><h4>4) LangGraph</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/300/1*xsWdK3_DGIAGApL00aDxAw.jpeg" /></figure><p>LangGraph is a promising framework for creating stateful, multi-actor applications with LLMs, built on LangChain. It extends LangChain Expression Language to coordinate multiple chains (or actors) across computation steps in a cyclic manner, inspired by Pregel and Apache Beam. LangGraph benefits from a robust community and the LangChain ecosystem, offering a powerful platform for developing multi-agent systems.</p><h3>Practical Applications of Multi-Agent Systems</h3><h4>1) Content Creation and Editing</h4><p>Multi-agent systems can also be applied to content creation and editing. For example, one agent could generate initial content, while another agent edits and refines it. Through collaborative interactions, these agents can produce polished and well-structured content, suitable for various purposes such as marketing, technical documentation, or creative writing.</p><h4>2) Code Generation and Review</h4><p>One practical application of multi-agent systems is in the field of code generation and review. Imagine a scenario where one agent generates application code while another reviews it. These agents can engage in iterative interactions to refine the code, producing higher quality results and reducing errors.</p><h4>3) Data Analysis and Insights</h4><p>In the realm of data analysis, multi-agent systems can provide significant benefits. Different agents can specialize in various aspects of data processing, analysis, and visualization. By working together, these agents can uncover insights, identify trends, and generate comprehensive reports. This collaborative approach can enhance the accuracy and depth of data analysis, leading to more informed decision-making.</p><h4>4) Customer Support and Service</h4><p>Multi-agent systems can revolutionize customer support and service by providing more responsive and personalized interactions. Different agents can handle various aspects of customer inquiries, from initial triage to detailed problem-solving. By coordinating their efforts, these agents can provide timely and effective support, improving customer satisfaction and loyalty.</p><h3>Content Creation and Editing using Crewai</h3><h4># Building Blocks of CrewAI</h4><p>CrewAI simplifies the integration of AI agents by breaking down the process into three main components:</p><ol><li><strong>Agent</strong>: The AI entities that perform specific tasks.</li><li><strong>Task</strong>: The individual assignments or objectives given to the agents.</li><li><strong>Crew</strong>: The collective group of agents working together on a project.</li></ol><h4>Example: Research and Write an Article</h4><ol><li><strong>Task</strong>: Research and write an article on a given topic.</li></ol><ul><li><strong>Description</strong>: A detailed explanation of the task to be performed.</li><li><strong>Expected Output</strong>: The desired result or product of the task.</li><li><strong>Assigned Agent</strong>: The specific agent responsible for executing the task.</li></ul><p><strong>2 . Agent Roles</strong>:</p><ul><li><strong>Researcher</strong>: Gathers data and insights.</li><li><strong>Content Writer</strong>: Structures the information and writes the article.</li><li><strong>Editor</strong>: Reviews and refines the final output.</li></ul><pre>import streamlit as st<br>from crewai import Agent, Task, Crew<br>from langchain_google_genai import ChatGoogleGenerativeAI<br>from dotenv import load_dotenv<br>import os<br>import nest_asyncio<br><br># Load environment variables<br>load_dotenv()<br><br># Apply nest_asyncio to allow nested event loops<br>nest_asyncio.apply()<br><br># Initialize Google Gemini AI<br>llm = ChatGoogleGenerativeAI(<br>    api_key=os.getenv(&#39;GOOGLE_API_KEY&#39;),<br>    model=&quot;models/gemini-pro&quot;  # Replace with the correct model name<br>)<br><br># Define Agents<br>planner = Agent(<br>    role=&quot;Content Planner&quot;,<br>    goal=&quot;Plan engaging and factually accurate content on {topic}&quot;,<br>    backstory=&quot;You&#39;re working on planning a blog article about the topic: {topic}. You collect information that helps the audience learn something and make informed decisions. Your work is the basis for the Content Writer to write an article on this topic.&quot;,<br>    llm=llm,<br>    allow_delegation=False,<br>    verbose=True<br>)<br><br>writer = Agent(<br>    role=&quot;Content Writer&quot;,<br>    goal=&quot;Write a compelling and well-structured blog post on {topic}.&quot;,<br>    backstory=&quot;You&#39;re a writer who uses the content plan to create a detailed blog post. Ensure it aligns with SEO best practices and the brand&#39;s voice.&quot;,<br>    llm=llm,<br>    allow_delegation=False,<br>    verbose=True<br>)<br><br>editor = Agent(<br>    role=&quot;Editor&quot;,<br>    goal=&quot;Edit a given blog post to align with the writing style of the organization.&quot;,<br>    backstory=&quot;You are an editor who receives a blog post from the Content Writer. Your goal is to review the blog post to ensure that it follows journalistic best practices, provides balanced viewpoints when providing opinions or assertions, and also avoids major controversial topics or opinions when possible.&quot;,<br>    llm=llm,<br>    allow_delegation=False,<br>    verbose=True<br>)<br><br># Define Tasks<br>plan = Task(<br>    description=(<br>        &quot;1. Prioritize the latest trends, key players, and noteworthy news on {topic}.\n&quot;<br>        &quot;2. Identify the target audience, considering their interests and pain points.\n&quot;<br>        &quot;3. Develop a detailed content outline including an introduction, key points, and a call to action.\n&quot;<br>        &quot;4. Include SEO keywords and relevant data or sources.\n&quot;<br>        &quot;5. The plan should be detailed and cover all aspects of the topic.&quot;<br>    ),<br>    expected_output=&quot;A comprehensive content plan document with an outline, audience analysis, SEO keywords, and resources.&quot;,<br>    agent=planner,<br>)<br><br>write = Task(<br>    description=(<br>        &quot;1. Use the content plan to craft a compelling blog post on {topic}.\n&quot;<br>        &quot;2. Incorporate SEO keywords naturally.\n&quot;<br>        &quot;3. Sections/Subtitles are properly named in an engaging manner.\n&quot;<br>        &quot;4. Ensure the post is structured with an engaging introduction, insightful body, and a summarizing conclusion.\n&quot;<br>        &quot;5. Proofread for grammatical errors and alignment with the brand&#39;s voice.\n&quot;<br>        &quot;6. Each section should have 2 or 3 paragraphs.\n&quot;<br>        &quot;7. The entire blog post should be at least 1000 words.&quot;<br>    ),<br>    expected_output=&quot;A well-written blog post in markdown format, ready for publication.&quot;,<br>    agent=writer,<br>)<br><br>edit = Task(<br>    description=(&quot;Proofread the given blog post for grammatical errors and alignment with the brand&#39;s voice.&quot;),<br>    expected_output=&quot;A polished and error-free blog post in markdown format, ready for publication.&quot;,<br>    agent=editor<br>)<br><br># Define Crew<br>crew = Crew(<br>    agents=[planner, writer, editor],<br>    tasks=[plan, write, edit],<br>    verbose=2<br>)<br><br># Streamlit App<br>def main():<br>    st.title(&quot;Content Creation Assistant&quot;)<br><br>    # Input for topic<br>    topic = st.text_input(&quot;Enter the topic for the blog post:&quot;)<br><br>    if st.button(&quot;Generate Content&quot;):<br>        if topic:<br>            with st.spinner(&#39;Generating content...&#39;):<br>                try:<br>                    result = crew.kickoff(inputs={&quot;topic&quot;: topic})<br>                    st.markdown(result)<br>                except Exception as e:<br>                    st.error(f&quot;An error occurred: {e}&quot;)<br>        else:<br>            st.error(&quot;Please enter a topic.&quot;)<br><br>if __name__ == &quot;__main__&quot;:<br>    main()</pre><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*YMYjCmBI7-8RxROSsUgrHA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*YuZ3AI7T3qyE5AU2w0XAMw.png" /></figure><p><a href="https://huggingface.co/spaces/Prathamesh1420/crewaigent">https://huggingface.co/spaces/Prathamesh1420/crewaigent</a></p><h3>Future Directions and Challenges</h3><h4># Evolving Frameworks</h4><p>The field of multi-agent systems is rapidly evolving, with new frameworks and tools continually being developed. Staying updated with the latest advancements and experimenting with different frameworks can help developers leverage the full potential of multi-agent systems.</p><h4># Addressing Ethical and Bias Concerns</h4><p>As with any AI system, ethical considerations and bias mitigation are essential. Multi-agent systems must be designed with fairness, transparency, and accountability in mind. Addressing biases in training data, ensuring diverse perspectives, and implementing robust evaluation mechanisms are critical steps in developing ethical and reliable multi-agent systems.</p><h3>Conclusion</h3><p>In summary, multi-agent architectures enhance collaboration among independent, language model-powered agents, resulting in more effective and modular solutions. By leveraging the strengths of multiple agents, these systems can tackle complex problems, improve efficiency, and scale to meet diverse needs. Several frameworks, such as OpenAI Assistant, Autogen, Dragonscale, CrewAI, and LangGraph, provide robust platforms for building and managing multi-agent applications.</p><h3>References</h3><ul><li><a href="https://microsoft.github.io/autogen/">AutoGen | AutoGen (microsoft.github.io)</a>.</li><li><a href="https://www.crewai.io/">crewAI — Multi AI Agents systems.</a></li><li><a href="https://python.langchain.com/docs/langgraph">🦜🕸️LangGraph | 🦜️🔗 Langchain</a>.</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=aaefa5350012" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>