Community workshop on Squad — feedback welcome 🙏

[joslat]

Hi all 👋

I've been building a hands-on workshop around Squad and wanted to share it with the community for feedback before polishing further: https://github.com/joslat/squad-workshop

Quick context on me: Microsoft AI MVP based in Zürich, working day-to-day with .NET / Semantic Kernel / Microsoft Agent Framework, and active in the Swiss .NET community. I also created & maintain AgentEval — a .NET-native memory Eval & benchmark suite for AI agents — which is partly why Squad's persistent-team-memory model caught my attention in the first place.

What it is
Three modules taking someone from zero to a real full-stack app driven by the team:

• Module 1 — Basic (~90 min): scaffold the team, plan, record a decision, build a vertical-slice .NET 10 + React reading-list app, run a reviewer pass.
• Module 2 — Intermediate (~45 min): second-wave feature with regression-aware tests, then inspect what the team actually left behind in .squad/ (the compounding-memory check).
• Module 3 — Advanced (~60 min, work in progress): Aspire observability + Ralph autonomous mode.

Honest scope
Modules 1 and 2 run end-to-end. Module 3 is still rough — I haven't fully nailed the "agentic team loop iterating over the mega-plan" yet, so right now it reads more as a guided tour than a build exercise. Polishing it is the next push.

What I'd love feedback on

• Does the framing match how you'd want newcomers to meet Squad?
• Anywhere the workshop oversells, undersells, or misrepresents how Squad actually behaves in practice?
• Module 3: any pointers on getting the autonomous loop reliable enough for a workshop setting?
• Teaching moments I might be missing — patterns or "aha" gates you'd want a learner to hit on the first run?

A related question, since people here might be thinking about it
Squad's central promise is that team memory compounds — the second feature is faster because the repo remembers. The question I keep wrestling with on the AgentEval side is: how do you actually measure that? How does a user verify their .squad/ state is genuinely compounding vs. quietly accumulating noise? If anyone has been thinking about memory diagnostics or eval against Squad teams, I'd love to compare notes — feels like a natural frontier and I haven't seen it explored publicly yet.

Happy to iterate on any of this. And if at some point it ever feels worth featuring or folding into the official org, I'm very open to that — but no expectation, the workshop stands on its own either way.

Thanks for building Squad — the team-as-files model has been genuinely fun to explore.

— José