Agent Teams: durable specialists, owned work, and coordination

[wSedlacek]

Summary

This discussion asks for a direction decision. Should T3 Code invest in Agent Teams as a product bet — durable specialist seats on a project, owned work with closures, review/attention handoffs, and coordination tools — beyond one-off delegate_task children and provider-native teammates?

This is larger than observing Claude’s native teammates (#33132). #34683 and related slices were a prior stack; Macroscope skipped the umbrella on review cost.

Why durable teams

Users already orchestrate multi-agent work ad hoc: spawn children, copy context, babysit handoffs in chat. That works for a afternoon and falls apart across restarts, remotes, and “who owns this piece?”

A durable team model makes long-running projects legible: named specialist seats, owned work that survives process death, and an explicit review path when work is ready for attention. Ephemeral delegation stays for throwaway tasks; provider-native teammates stay as observation (#33132). T3-owned roster is the missing middle.

Proposal

Split along user-visible seams, each with its own prior approval:

1. Observe provider-native teammates in the shared Agents UI (small, provider-shaped).
2. Durable specialist roster on a project.
3. Owned work + closures / work_review.
4. Workshop / contributor bundles.

Remote seat addressing and next-turn provider switch are later layers, not MVP. Do not ship (1)–(4) as one portfolio PR.

Reference implementation

#34683 (Agent Teams umbrella) plus specialist / owned-work / work_review / bundle slices under that effort. Linked as evidence of demand and of review-budget failure modes — prefer the split above over reopening the umbrella.

Related: #33132 (observe native Claude teammates) — complementary observation, not T3-owned seats.

Question for maintainers

Yes, no, or not now: T3-owned durable specialist teams (roster → owned work → review), shipped in separately approved slices — not only ephemeral delegate_task + provider-native teams.

[wSedlacek]

Note

🤖 Composer responding on behalf of wSedlacek

Prior effort: #34683 (umbrella) plus specialist/owned-work slices such as #34618, #34591, #34593.

What we learned trying it

• Durable seats + owned work + review handoffs are compelling, but shipping them as one portfolio exceeds review budget (Macroscope skipped the umbrella on cost).
• Split along user-visible seams: (1) observe provider-native teammates, (2) durable specialist roster, (3) owned work/closures, (4) workshop/bundles. Each needs its own prior approval.
• Remote seat addressing and next-turn provider switch are powerful and also easy to get wrong across environments — treat as later layers, not MVP.