Coordinator/Root Copilot takes over?

[jamescarter-le]

I've been experimenting with Squads with generally good results, however when I'm working in an interactive way with Copilot and the Squads, sometimes it tends to 'break out' of the Squads and the root copilot AI or agent starts debugging and correcting things, not passing them to the appropriate 'team member'. This is usually happening when tests are failing and it gets tunnel vision on a task, or I'm giving hints/advice to the system.

Additionally, it would be nice to be able to give some input to an agent while it's running, if it's missing a piece of information that is obvious to me, but it will take the agents a lot time to determine something, or I know an invariant about a dependency of the system, it won't pick this up until the 'task' is finished. How can I 'inject' some knowledge to the team while they are running?

Thanks! Still experimenting and enjoying Squads.

[bradygaster]

This does happen quite often; we've tried many things but it seems the coordinator truly loves to control the situation. In the upcoming "big release" right before 1.0, we'll have a new CLI that you can use if you really want to 86 the Copilot coordinator. It will mean you'd sacrifice all the other GitHub Copilot CLI features whilst you're "squadding."

I typically solve this by explicitly using terms like "team fan out," which is a sure-fire way to make sure the team does the work. Or, direct the request to the agent you need or, to that agent "and the team." All of these prompt techniques will help push Copilot to push the request to the squad.

[jamescarter-le]

Understood I also use similar keywords to getting it back on track. Looking forward to seeing how Squad evolves, it's been the best experience to develop applications with a lot of complexity and cross cutting concerns where most AI fails and gets confused by the end.

[bradygaster]

Thanks for exploring Squad and for the thoughtful feedback, @jamescarter-le!

This is a known behavior we're aware of. Some workarounds that help: use phrases like "team fan out" or direct requests to specific agents (e.g., "Fenster, look at the error handling"). These cues push toward delegation.

We're also working on a dedicated CLI experience that gives you more direct control over the squad orchestration. Stay tuned for updates! 🚀

[bradygaster]

📋 Filed a tracking issue to improve coordinator delegation behavior. This is a real UX concern and we're working on it — the CLI-First architecture (#252) and skill-based orchestration (#255) should both help significantly. Thanks for the detailed feedback, @jamescarter-le!

Linked issue: #257