🤖 Microsoft Build 2026: 7 Moves That Matter

Microsoft just ran Build 2026, and the announcements land differently if you ship with AI for a living.

Scout is in Teams now, a persistent AI coworker that handles office tasks without clocking out. MAI-Thinking-1 (35B reasoning) and MAI-Code-1-Flash (5B, GitHub Copilot-native) are out for early partners. GitHub's Kyle Daigle spelled out the platform's agent roadmap. And the Economist asked a quiet question the industry keeps dodging: can public markets actually absorb Anthropic, SpaceX, and OpenAI at these valuations?

Today we have 5 signals, 6 repos, 3 tools, and a tease of Friday's builder kit, a tool that reads the contract clause your client is hoping you'll never notice.

1. Microsoft Build 2026 led with agents. Scout (a persistent Teams agent), two new MAI models, and a refreshed Copilot stack were the keynote spine. The Verge has the full rundown.

2. Microsoft's MAI-Thinking-1 and MAI-Code-1-Flash are shipping. The 35B reasoning model goes to select early partners; the 5B code model targets GitHub Copilot directly. Both signal Microsoft building model muscle independent of OpenAI. Simon Willison has the technical breakdown.

3. GitHub published its agent roadmap. Kyle Daigle's Latent Space interview covers how the platform is handling the strain of agentic coding at scale, and what comes next for Copilot. Latent Space

4. The Economist asked if public markets can swallow Anthropic, SpaceX, and OpenAI. The valuations are enormous; the IPO window is tightening. Worth reading before Friday's kit. The Economist

5. Travelers deployed an OpenAI-powered claim assistant nationwide. It handles 24/7 customer guidance and scales automatically during catastrophe spikes, a clean production case study for insurance AI. OpenAI Blog

[Repo] chopratejas/headroom, Compresses tool outputs, logs, files, and RAG chunks before they hit the LLM. Claims 60, 95% token reduction with no meaningful accuracy loss. Ships as a library, a proxy, and an MCP server, you pick the integration surface.

[Repo] colbymchenry/codegraph, Pre-indexed code knowledge graph for Claude Code, Codex, Gemini, and Cursor. Fewer tokens and fewer tool calls because the graph answers structural questions locally before the model ever tries.

[Repo] Lum1104/Understand-Anything, Turns any codebase into an interactive knowledge graph you can explore, search, and question. Works with Claude Code, Cursor, Copilot, and Gemini CLI, useful when you inherit a repo and need to understand it fast.

[Repo] affaan-m/ECC, Agent harness optimization system covering skills, instincts, memory, and security for Claude Code, Codex, Opencode, and Cursor. Built as a research-first performance layer for operators who want more predictable agent behavior.

[Repo] microsoft/markitdown, Python tool that converts Office documents, PDFs, and files into clean Markdown. If your pipeline ingests anything outside plain text, this is the pre-processing step you probably need.

[Repo] langgenius/dify, Open-source LLM app platform with a drag-and-drop RAG pipeline builder, workflow editor, and self-host option. Production-ready with an active community; useful if you want a full agent app without assembling the stack yourself.

[Tool] repomix, Packs an entire repo into a single LLM-ready file (XML, Markdown, or plain text). The non-obvious move: run it before every Claude session on a large codebase so the model gets the whole picture in one context load instead of hunting across files.

[Tool] simonw/llm, Simon Willison's CLI and Python library for running prompts against any model, including local ones. The plugin ecosystem is the value: you can swap models, log outputs, and chain calls without rewriting your scripts.

[Tool] paul-gauthier/aider, Terminal-based AI pair programmer that's git-aware and handles multi-file edits natively. Architect mode lets you plan a refactor in plain English and then execute it across the full codebase in one session.

Project memory with CLAUDE.md, teach Claude what your project does once, and it remembers every session.

Claude Code reads a file called CLAUDE.md automatically every time you start a session in a project folder. Think of it as a standing briefing note: tech stack, coding conventions, what not to touch, and anything the model should know before it types a single character.

1. In your project root, create the file: touch CLAUDE.md 2. Open it and add a few lines in plain English. For example:

# Project context Stack: Next.js, Supabase, Stripe Run tests before every commit: npm test Never edit files in /legacy, read-only

3. Start a Claude Code session in the same folder. It will read that file before anything else.

You'll know it worked when Claude references your stack or your rules without you mentioning them, try asking "what stack are we using?" right at the start of a fresh session.

Claude Code docs: memory

The Economist's question, can markets absorb Anthropic, SpaceX, and OpenAI, is really a question about whether AI infrastructure compounds like software or burns like infrastructure. Stratechery's piece on Google issuing equity to Berkshire Hathaway points the same direction: capital is becoming the durable moat, not code. For operators, that means the window to build on top of these models before pricing shifts is shorter than it looks.

This Friday's kit is ClauseGuard, and the spine of it is one uncomfortable fact: the clause that lets your client keep your finished work without paying you is probably sitting on page 4, written in language designed to be skipped. Most freelancers and small agencies sign contracts they don't fully understand because a $400/hr lawyer isn't in the budget. Claude reads dense legal language, reasons about downstream risk, and drafts counter-clauses, which means this is now a $19/month product, not a $400/hr conversation. Full build, stack, and revenue math drop Friday. Get early access

See you tomorrow.