How can we get claude agent teams involved into agor?

[vedantprajapati]

Claude agent teams came out and it has blown my expectations with parallelizing complex tasks. How can we use it in agor?

[mistercrunch]

Haven't tried the feature yet, how does it work mechanically in Claude Code? Supported by the SDK?

[vedantprajapati]

https://code.claude.com/docs/en/agent-teams#how-agent-teams-work

https://kargarisaac.medium.com/agent-teams-with-claude-code-and-claude-agent-sdk-e7de4e0cb03e

I believe it's just an env var you need to enable with ClaudeAgentOptions.

Usually, Subagents only report results to the main agent and never talk to each other. But with agent teams, teammates share a task list, claim work, and communicate directly with each other. IIRC, this is done by spawning terninal sessions per teams agent.

[mistercrunch]

Yeah looks like they share a centrally managed todo list, wondering about the internals, whether the todo list is the only communication/interface. I see some optional tmux/iterm integrations where it spins off other terminals (optionally I guess). Wondering if setting the basic mode + env var would work within the SDK, worth a shot. I think personally I prefer the coordinator/supervisor model. Agor might be outdoing the this feature in some ways. Feature is marked as experimental I think, curious to see where that goes.

Totally possible to have similar workflows through deterministic or agent-managed in Agor. Landed https://agor.live/blog/agor-assistants today. Opens up variety of supervisor-type workflows and use cases, like "make a todo list or DAG of tasks, and constantly keep 12 agents busy working in parallel hammering at it". Wondering if you "spin off 5 agents, point them to a todo list you manage, and instruct the agents to keep working until all is done". I guess you'd need some instructions as to how to manage/update the central todo list.

I really need to try the feature directly in the CLI and see whether/how it's worthwhile and take inspiration. My mindset lately has been more oriented towards supervisor/delegator-type models.

A bit worried about chain-reaction type things. Haven't seen much signs of it, but triggering agents that trigger agents that trigger other agent can go crazy. A bad line in CLAUDE.md saying "prefer delegation over doing work, systematically delegate to an agent whenever possible" could create an infinite chain :)

[vedantprajapati]

Yeah, you'd have to weigh the pros and cons with it, it's currently a little slow and still is buggy. With agor, you can kind of tell which subsessions should feed back information to the parent.

With this, it's just allowed. The granular control isn't there just yet, but I suppose teams is something anthropic is actively working on.

For now, probably a good idea to just track this mentally until it gets polished.

[mistercrunch]

My take: at some point managing groups of agents becomes a UX problem, at least if you want observability/control. Only so much they can do with terminal windows.

About "feedback to parent", I think it could be great to enable/disable "notify parent on prompt completion" more actively in UI, probably belongs close to the prompt box, and would be a simple toggle. I spawn subsessions a lot for code review, but don't always want to callback the parent throughout a session.