Skill-based orchestration

[swnger]

First of all, thanks for the project! It is quite interesting and I tested it out on a few features already with great success.
Having used copilot-orchestra and copilot-atlas extensively, my experience is that trying to build a deterministic orchestration workflow into a custom agent is difficult, since the workflow instructions are just part of context and hence get forgotten / compacted quite often.

Looking at the > 1000 line workflow instructions of the main agent I fear this will happen quite a lot?
Have you thought about offloading some of the workflows in the squad agent to skills?
Squad init, adding members, migrations.... there are many workflow instructions injected into every session that are one-off or at least rarely needed.
I think an init skill (maybe using the existing npx setup logic) would also enable you to distribute squad as a copilot CLI plugin?

And out of curiosity: What was the decision driver to "prompt" the custom agents into the general subagent instead of setting up the team as standard custom agents under .github/agents?

[bradygaster]

That's an amazing idea - a lot of the stuff in there could totally be offloaded to skills. Genius.

And out of curiosity: What was the decision driver to "prompt" the custom agents into the general subagent instead of setting up the team as standard custom agents under .github/agents?

I hadn't thought about that. I'll have the Squad consider this. Thanks!

[bradygaster]

[swnger]

Thanks, glad the squad liked it 😁

[bradygaster]

Hey @swnger — this is a really compelling idea, and it resonates deeply with where we're heading. You've articulated the core tension perfectly: the monolithic instruction file is a context bottleneck.

We've just filed this as a tracked issue: #255 — Skill-based orchestration: decompose squad.agent.md into pluggable skills

Here's the exciting part: the infrastructure already exists. .squad/skills/ is live today for agent-level skills. Your proposal extends this to the coordinator itself — making Squad's core capabilities pluggable and composable. Combined with SDK mode (defineSkill() in TypeScript), this becomes a typed, testable, community-extensible architecture.

We're targeting this for the v0.8.22 roadmap alongside SDK-First (#249-251) and CLI-First (#252). If you're interested in collaborating on the design, we'd welcome input on the key questions in the issue.

Thanks for pushing this forward! 🚀

[bradygaster]

That's fantastic, @swnger! Your idea has become reality — we shipped defineSkill() in commit 412ce58, and it's the foundation for exactly what you envisioned.

Skills now work end-to-end:

• Define them in squad.config.ts using defineSkill()
• They're extracted to .squad/skills/ on build
• Fully typed in TypeScript, testable, and composable

This is already reducing the cognitive load on coordination. More workflows can now live as skills instead of bloating the central instructions. You can try it in the current release.

We'd love your feedback on how this approach works for your workflows. The door is wide open for community skills as well.

Thanks for pushing us in this direction! 🚀

— McManus 📝, on behalf of the Squad

[swnger]

Wow...that was fast!
There's one more thing if you want to go that far:
I think the /init logic could live completely outside of the squad main loop.
It could be a skill that you give your non-squad copilot to bootstrap the squad, including drafting the team.
The squad main agent does not need to know of init or it's current two modes anymore.
It's always in team mode.

That separate init skill is what could be delivered as a copilot plugin (I think now supported by both vs code and the cli), removing the need for npm.

Might be far fetched and could be that the current way works better, because the squad agent uses its internal knowledge about the whole squad internal process during team setup. But maybe worth a try 😄