Squad, an open source project from Brady Gaster built on GitHub Copilot, deploys a preconfigured multi-agent AI team directly inside your repository with two commands: `npm install -g @bradygaster/squad-cli` globally, then `squad init` per repo. You get four specialists: a lead, frontend developer, backend developer, and tester. No vector databases, no orchestration framework, no prompt engineering prerequisites.
The architecture makes three specific bets worth understanding. First, shared memory lives in a versioned `decisions.md` file, not a live session or real-time sync layer. Second, the coordinator agent is a thin router only: each specialist runs as a separate inference call with up to 200K tokens of context, so you replicate repository context across agents rather than split one window among four. Third, reviewer automation blocks the original author agent from fixing its own rejected code, forcing a separate agent with a fresh context window to step in. That last point is the most technically interesting part of the piece.
Squad is not autonomous. Agents ask clarifying questions, make wrong assumptions, and every pull request still requires human review and merge. What the full article details is how the `.squad/` folder structures agent identity through plain-text charter and history files that get cloned with the repo, meaning collaborators inherit an already-onboarded AI team without any setup. That mechanism, versioned agent memory as a first-class repository artifact, is the design pattern worth reading for.
[READ ORIGINAL →]