Microsoft's Multi-Agent AI System

Today’s top AI Highlights:

1. Microsoft’s opensource framework to orchestrate a team of AI agents for general tasks

2. Build, monitor, and debug AI agents in real-time with a visual interface

3. Gemini is now accessible via the OpenAI libraries and REST API

4. Cohere opensources cohere-finetune for high-quality fine-tuning of Cohere's models

5. Fast GraphRAG tool for cost-efficient, high-precision, agent-driven RAG workflows

& so much more!

Read time: 3 mins

AI Tutorials

xAI API is finally here with the new grok-beta model. This model comes with 128k token context and function calling support. Till 2024 end, you even have $25 free credit per month!

We just couldn’t resist building something with this model so here it is! We are building an AI Finance Agent that provide current stock prices, fundamental data, analyst recommendations for stocks, and search for the latest news of a company to give a holistic picture.

It uses:

• xAI's Grok for analyzing and summarizing information

• Phidata for agent orchestration

• YFinance for real-time stock data from

• DuckDuckGo for web search

We share hands-on tutorials like this 2-3 times a week, designed to help you stay ahead in the world of AI. If you're serious about levelling up your AI skills and staying ahead of the curve, subscribe now and be the first to access our latest tutorials.

Don’t forget to share this newsletter on your social channels and tag Unwind AI (X, LinkedIn, Threads, Facebook) to support us!

Latest Developments

Build Custom Multi-Agent Tools for Daily Tasks 👫

Microsoft has released an opensource multi-agent generalist system Magentic-One that goes beyond areas like software engineering, data analysis, and web navigation, to solve complex, open-ended tasks like web and file operations, much like everyday work tasks that people encounter in their daily lives.

Built on the very famous AutoGen framework, Magentic-One employs a unique architecture where a central "Orchestrator" agent coordinates the activities of specialized agents like WebSurfer, FileSurfer, Coder, and Computer Terminal. This modular design helps to extend and adapt the system's functionalities easily.

Key Highlights:

1. Modular system - The system offers a plug-and-play architecture where you can swap or add agents without touching the core system. Each agent maintains its own state and communication channels. The system includes built-in error handling and recovery mechanisms.

2. Monitoring - The framework provides comprehensive logging and monitoring tools that track every agent action, decision, and state change. This includes browser interactions, file operations, code execution, and inter-agent communications to debug workflows.

3. Production-ready features - The framework includes features like rate limiting and error handling. Security features include containerization, human-in-the-loop options, and clear audit trails of agent actions.

4. Built-in Evaluation - AutoGenBench, an open-source evaluation tool, comes bundled with Magentic-One to measure agent performance, test system behavior, and validate custom implementations. It includes pre-built test scenarios, isolation controls to prevent unintended side effects, and detailed performance metrics.

5. Flexible LLM Integration - While it defaults to GPT-4o, Magentic-One is model-agnostic. You can swap in different LLMs and specialized LMs to optimize for performance, cost, or specific capabilities.

Make Unwind AI work better for you!

Over the past few months, we've noticed our community has some seriously impressive developers building cool stuff with AI. We'd love to know more about what you're working on so we can make Unwind AI even more useful for your specific needs.

Could you spare 2 minutes for 6 questions?? Our super-quick survey will help us:

• Focus on the kind of tools and frameworks you actually use

• Create more relevant content

• Share the kind of tutorials and guides you’d want to see

→ Take the 2-min survey

Your input will directly shape our upcoming content. Promise it's just 6 quick questions – no essay writing required!

Appreciate your help in making Unwind AI more valuable for your daily work.

Watch How Your AI Agent Thinks 🔮

Burr is an opensource Python library to build and debug AI agents using state machines. The library provides a visual interface to track how your AI agents make decisions, showing each step of the logic flow in real-time through an interactive graph. Burr eliminates the need to add extensive logging or debug prints to understand agent behavior.

Burr addresses common pain points in building AI agents - managing conversation history, handling complex decision trees, and debugging unexpected outputs. If you’re building anything with complex logic, like chatbots, agents, or simulations, Burr is worth checking out.

Key Highlights:

1. Rapid development - Create AI agents using simple Python functions with clear inputs/outputs. No need to learn complex abstractions - if you can write a Python function, you can build an agent. Integrates with OpenAI, Anthropic, and other LLM providers.

2. Visual debugging interface - Watch your agent's decision-making process through a graph UI that updates in real-time. The built-in monitoring interface shows you execution traces and lets you click on any node to inspect the exact state and data at that point.

3. Production-ready features - Built-in support for streaming responses, state persistence with SQLite/PostgreSQL, and async APIs for web services. The state machine approach makes it easy to add safety checks and fallback behaviors.

4. Framework-agnostic - Works with any Python code, not just LLMs. Build agents that combine language models, traditional ML, business logic, and external APIs. Add custom actions and state handlers without touching the core library.

5. Better than LangGraph - While both frameworks handle state machines, Burr brings an opensource monitoring UI, works with non-LLM applications, and provides more extensive telemetry. Unlike LangGraph, it comes with out-of-the-box persistence options and a complete toolkit for building, debugging, and deploying state-aware apps.

Quick Bites

Gemini models are now accessible through the OpenAI Library and REST API, supporting Chat Completions and Embeddings APIs for quick integration. You can start experimenting with Gemini right away—check out the example code provided to test the Chat Completions feature in Python, JavaScript, and REST.

Building a top-tier code LLM is no longer a secret recipe. OpenCoder is a fully opensource coding LLM that matches top-tier performance across major benchmarks. The team has shared its complete training data pipeline, reproducible 2.5-trillion-token dataset, and detailed training protocols. Beyond just model weights, you now have access to OpenCoder's entire "cookbook" - including data cleaning rules, deduplication methods, and instruction tuning strategies - so you can understand, reproduce, and build a high-performing code LLM from scratch.

Cohere’s new GitHub repo "cohere-finetune" lets you easily fine-tune Cohere's models like Command R and Aya Expanse with LoRA and QLoRA, using your own data for specific applications. The repo supports various base models and fine-tuning strategies, providing Dockerized setup for experimentation and deployment. It requires at least one GPU (with specific memory requirements varying based on the chosen model, batch size, and sequence length), though they have showcased optimal performance examples using 8x H100s.

Tools of the Trade

1. Fast GraphRAG: A high-speed, cost-efficient version of GraphRAG, offering a 6x cost reduction and faster processing for large-scale RAG pipelines. It enhances retrieval workflows with features like real-time updates, dynamic graph generation, and intelligent, interpretable data exploration.

2. ChainForge: Opensource visual programming environment for prompt engineering with minimal coding. It lets you evaluate, visualize, and refine prompt performance by setting up metrics and comparing responses from different LLMs and settings.

3. Parse Partial JSON Stream: Converts a stream of tokens into JSON format in real-time, so your app can show results as they load instead of waiting until the end. It helps make AI apps feel faster and more interactive for users.

4. Awesome LLM Apps: Build awesome LLM apps using RAG to interact with data sources like GitHub, Gmail, PDFs, and YouTube videos with simple text prompts. These apps will let you retrieve information, engage in chat, and extract insights directly from content on these platforms.

Hot Takes

1. sam altman: we are a few 1000 days away from building god. we will build suns on earth, unify physics and resurrect the worthy dead
   garry tan: sounds like this will be really impactful for startups
   sam altman: definitely. no better time to be a startup ~
   typedfemale

   
   

2. the fact that everyone thinks openai has lost all its talent everytime one person leaves just means that people don’t have a good model of sheerly how much talent is at openai ~
   Rohan Pandey

That’s all for today! See you tomorrow with more such AI-filled content.

Don’t forget to share this newsletter on your social channels and tag Unwind AI to support us!

Unwind AI - X | LinkedIn | Threads | Facebook

Awesome LLM Apps | Sponsor Us

PS: We curate this AI newsletter every day for FREE, your support is what keeps us going. If you find value in what you read, share it with at least one, two (or 20) of your friends 😉