Microsoft’s Deep Research Agent for Large Codebases

Today’s top AI Highlights:

1. The first Deep Research agent for large systems code

2. Connect any React app to any MCP server in just 3 lines of code

3. General-purpose AI agent for multi-step tasks with native MCP integration

4. Chat with your database like it's your most experienced teammate

5. Drop a URL, get a RAG API back instantly with zero setup

& so much more!

Read time: 3 mins

AI Tutorial

We've been stuck in text-based AI interfaces for too long. Sure, they work, but they're not the most natural way humans communicate. Now, with OpenAI Agents SDK and their text-to-speech models, we can build voice applications without drowning in complexity or code.

In this tutorial, we'll build a Multi-agent Voice RAG system that speaks its answers aloud. We'll create a multi-agent workflow where specialized AI agents handle different parts of the process - one agent focuses on processing documentation content, another optimizes responses for natural speech, and OpenAI's text-to-speech model delivers the answer in a human-like voice.

Our RAG app uses OpenAI Agents SDK to create and orchestrate these agents that handle different stages of the workflow.

We share hands-on tutorials like this every week, designed to help you stay ahead in the world of AI. If you're serious about leveling 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) to support us!

Latest Developments

Microsoft’s Deep Research Agent for Large Codebases 🔍 💻

AI coding agents have gotten pretty good at building apps, fixing GitHub issues, and handling pull requests on typical codebases.

But when it comes to massive systems like the Linux kernel with 75,000 files, 28 million lines of code, and decades of development history, even the best agents fall flat.

Microsoft just released Code Researcher, the first deep research agent designed specifically for these complex system-level challenges. The agent doesn't just read crash reports; it conducts multi-step reasoning over code definitions, tracing data flows, hunting for patterns, and diving deep into commit histories to understand how bugs evolved over time.

Key Highlights:

1. Deep codebase exploration - Code Researcher explores an average of 10 files per trajectory compared to just 1.33 files for SWE-agent, using specialized reasoning strategies to chase control flows, identify patterns, and analyze historical commits.

2. Outperforms existing agents - Using GPT-4o and o1 as LLMs, it resolves 58% of Linux kernel crashes versus 37.5% for SWE-agent, with significantly better overlap between generated patches and developer-referenced context.

3. Historical commit analysis - First coding agent to leverage commit history for bug fixing, with ablation studies showing a 10% performance drop when this capability is removed.

4. Structured context memory - Maintains organized memory of search results and reasoning steps, filtering relevant information during patch synthesis to avoid getting lost in irrelevant details.

The paper is currently under academic review as a preprint but the implementation, dataset, prompts, and evaluations should be opensourced soon.

Plug Any React App into MCP Servers Instantly ⚛️🪝

Building MCP clients just got stupid simple.

Cloudflare dropped an opensource React library that connects your app to any MCP server in literally 3 lines of code - no more wrestling with transport protocols, auth flows, or session management. The use-mcp library handles all the messy backend stuff automatically.

They have also opensourced AI Playground that gives you a complete chat interface that works with multiple LLMs and MCP servers right out of the box. You can deploy, customize, and extend both the tools however you need for your AI applications.

Key Highlights:

1. Instant MCP integration - Drop the useMCP() hook into any React app and you're connected to remote MCP servers with automatic tool discovery, connection retries, and real-time state management.

2. Complete auth handling - Built-in OAuth 2.1 support manages the entire authentication flow, token storage, and credential management without writing any custom logic.

3. AI Playground - Open-source chat interface with Workers AI integration, MCP server connections, and debugging tools that you can deploy and customize immediately.

4. Future-proof architecture - Supports both Server-Sent Events and Streamable HTTP transports with automatic protocol detection and backwards compatibility as MCP standards evolve.

Quick Bites

Midjourney just dropped their first video model, and it's refreshingly honest about being a "stepping stone" rather than the final destination. The V1 model transforms static images into four 5-second clips through Discord, offering both automatic animation and manual motion prompts with high or low motion settings. At roughly 8x the cost of image generation, it's still 25x cheaper than existing market options.

After dropping their impressive M1 reasoning model, China’s MiniMax didn't pause for applause. They followed with MiniMax Agent, a general-purpose AI agent built for long-horizon tasks with native MCP integrations. The agent handles multi-step planning, task decomposition, and end-to-end execution with built-in connections to GitHub, Slack, and Figma, plus native multimodal generation capabilities. It’s completely free to try with generous credits.

MiniMax frames it perfectly: in the age of intelligent agents, "code is cheap, show me the requirement."

Tools of the Trade

1. MongoDB MCP Server: Stop explaining your database schema to AI – let it connect directly and see for itself. This MCP server gives AI assistants native MongoDB access, letting you run aggregations, manage clusters, and perform CRUD operations directly through chat.

2. Firestarter: Opensource tool to instantly create a knowledgeable AI chatbot for any website. Firestarter crawls your site, indexes the content, and provides a ready-to-use RAG-powered chat interface and an OpenAI-compatible API endpoint.

3. Claude Code Usage Monitor: A beautiful real-time terminal monitoring tool for Claude AI token usage. Track your token consumption, burn rate, and get predictions about when you'll run out of tokens. Auto-updates every 3 seconds.

4. Awesome LLM Apps: Build awesome LLM apps with RAG, AI agents, MCP, and more to interact with data sources like GitHub, Gmail, PDFs, and YouTube videos, and automate complex work.

Hot Takes

1. I really like the term “context engineering” over prompt engineering.

   It describes the core skill better: the art of providing all the context for the task to be plausibly solvable by the LLM.~
   Tobi Lutke
   
   

2. Meta hiring spree highlights that the scarcest resource in AI is not technical talent, but leadership with taste, just as models themselves are more taste-oriented

   The core algorithms of sota models are pretty much the same, the data & how it is used differs

   Really rare talent ~
   Emad Mostaque

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

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 😉