agent-teams

# Agent Teams > **Experimental**: Agent teams require `CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1` and may change between Claude Code versions. > **BREAKING (Claude Code 2.1.178)**: The explicit `TeamCreate` / `TeamDelete` tools were **removed**. Every session now has **one implicit team** when `CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1` is set — there is nothing to create or tear down. The `team_name` parameter is still **accepted but ignored** on the tools that took it (e.g. `Agent`), so older invocations don't error; omit it in new code. Coordination is otherwise unchanged: spawn teammates with the `Agent` tool, coordinate with `SendMessage`, and track work with `TaskCreate`/`TaskList`/`TaskUpdate`. > **Task tools are opt-in on current models (Claude Code 2.1.233+).** `TaskCreate` / `TaskGet` / `TaskUpdate` / `TaskList` / `TodoWrite` are not exposed on Opus 4.8, Sonnet 5, Fable 5 / 5.1, Mythos 5 or newer unless the session opted in — see `.claude/rules/agentic-permissions.md` § "Standard Permission Sets" → "Task-tool availability" for the three opt-ins and the full semantics. On a session without one, a lead cannot create or assign shared tasks: set the variable, or coordinate via `SendMessage` alone and carry the task list in the lead's own message to each teammate. Check `TaskList` is callable before spawning teammates. For the worked setup examples, communication snippets, shutdown procedures, the worktree path-resolution recovery routine, and common dispatch patterns, see [REFERENCE.md](REFERENCE.md). ## When to Use This Skill | Use agent teams when... | Use subagents instead when... | |------------------------|------------------------------| | Multiple agents need to work in parallel | Tasks are sequential and interdependent | | Ongoing communication between agents is needed | One focused task produces one result | | Background tasks need progress reporting | Agent output feeds directly into next step | | Complex workflows benefit from task coordination | Simple, bounded, isolated execution | | Independent changes to the same codebase (with worktrees) | Context sharing is fine and efficient | ## Sub-Agent Caveat: Spawn Teams from the Main Thread `Agent` and the related parallel-spawn tools may not be present in a **sub-agent's** tool surface, even if the parent conversation has them. A sub-agent designed to orchestrate its own team can silently degrade to sequential single-thread execution — same content, ~5× longer wall-clock — without surfacing the failure until its post-completion summary. | Situation | Recommended pattern | |-----------|---------------------| | Fan-out from the main conversation | Spawn parallel `Agent` calls directly — full tool surface available | | Sub-agent orchestrating its own team | Avoid by design: split the work so the main thread does the fan-out | | Sub-agent must orchestrate a team | Detect tool availability up front; report sequential fallback as a first-class outcome | **Detection contract** to brief into a coordinating sub-agent: confirm the `Agent` tool is callable; if not, do **not** silently fall back — report "Parallel fan-out unavailable in this sandbox; executed sequentially." as the first line of the summary, then continue sequentially with the same input contract. Plan top-level orchestration in the main conversation when you can. ## Native Team Tools The implicit team (2.1.178) needs no create/delete step — these tools operate on the one team the session already has: | Tool | Purpose | |------|---------| | `Agent` | Spawn a teammate (give it a `name`; `team_name` is accepted but ignored) | | `SendMessage` | Send DMs, broadcasts, shutdown requests, plan approvals | | `TaskCreate` / `TaskList` / `TaskUpdate` | Create, discover, and assign/advance shared tasks | | `TaskOutput` | Get output from a background agent | | `TaskStop` | Stop a running background agent | The setup sequence — `TaskCreate` → spawn teammates via the `Agent` tool (with a `name`) → `TaskUpdate` to assign → receive results automatically — is shown with full code in [REFERENCE.md → Team setup workflow](REFERENCE.md#team-setup-workflow). Teammate messages are delivered to the lead's inbox between turns; no polling needed. ## Task Management | State | Meaning | |-------|---------| | `pending` | Not yet started | | `in_progress` | Assigned and active (one at a time per teammate) | | `completed` | Finished successfully | | `blocked` | Waiting on another task (`blocked_by` field) | Teammates claim tasks in **ID order** (lowest first) via `TaskList` + `TaskUpdate`, and skip blocked tasks until their blocker completes. ## Communication (SendMessage) | Type | Use When | |------|----------| | `message` | Direct message to a specific teammate (recipient = **name**, not agentId) | | `broadcast` | Critical team-wide announcement (N teammates = N round-trips — use sparingly) | | `shutdown_request` / `shutdown_response` | Graceful teammate exit handshake | | `plan_approval_response` | Approve or reject a teammate's plan | DM, broadcast, and discovery (`Read ~/.claude/teams/<name>/config.json` → `members`) examples are in [REFERENCE.md → Communication](REFERENCE.md#communication-sendmessage). ### Key Teammate Rules - Mark exactly ONE task `in_progress` at a time. - Use `TaskUpdate` (not `SendMessage`) to report task completion. - Idle after every turn is normal — idle ≠ unavailable; a message wakes a teammate. - All communication requires `SendMessage` — plain text output is NOT visible to the lead. - Use the **name** field (not agentId) for `recipient`. ## Lead Preflight Checklist Before drafting the PRP and launching agents, a 30-second sweep prevents multi-edit renaming work after agents return: | Check | Command | Why | |-------|---------|-----| | Next ADR/PRD/PRP sequence number | `ls docs/blueprint/adrs/ \| sort -V \| tail -1` | Prevents numbering collisions in parallel doc writes | | Filename conflicts | `git ls-files \| grep <filename>` | Scope tables can't guard against a stale mental model of the tree | | Hardware pin budget (embedded) | Read `pin_config.h` or equivalent | Prevents pin assignments overlapping across Phase 1 agents | ## Out-of-Scope Discovery Protocol Include the out-of-scope discovery protocol in every agent's prompt when that agent has an exclusive write scope — it prevents the "investigate out of scope → exhaust budget → truncated summary" failure mode. Copy the verbatim block from [REFERENCE.md → Out-of-scope discovery protocol](REFERENCE.md#out-of-scope-discovery-protocol): the agent **stops immediately** on an out-of-scope dependency, lists it under an `Out-of-scope dependencies` summary section, and exits for the lead to triage. ## Worktree Isolation Hazard Worktree-isolated agents can have `Edit`/`Write` silently resolve relative paths against the **parent repo** instead of their worktree (upstream bug [#1091](https://github.com/laurigates/claude-plugins/issues/1091)), landing commits on the wrong branch with no immediate signal. Harden every worktree-isolated prompt with the absolute-path preamble, run the lead post-flight check (`git diff origin/main..HEAD` + `git status --porcelain` from the parent), and use the cherry-pick + rebase recovery — all in [REFERENCE.md → Worktree path resolution](REFERENCE.md#worktree-isolated-editwrite-path-resolution-1091). See also `.claude/rules/agent-coworker-detection.md`. ## Resuming Worktree-Isolated Agents `SendMessage`-resume of a completed worktree-isolated agent (`Agent` with `isolation: "worktree"`) **loses worktree isolation** — the resumed run executes in the orchestrator's main checkout, not the worktree. For parallel file-mutating work, re-dispatch a fresh `isolation: "worktree"` agent instead of resuming; reserve `SendMessage`-resume for read-only / single-checkout continuations. See `parallel-agent-dispatch` → "Resuming agents: SendMessage loses worktree isolation" (issue [#1546](https://github.com/laurigates/claude-plugins/issues/1546)). ## Sandbox Considerations In web sessions (`CLAUDE_CODE_REMOTE=true`): - Sub-agents (teammates) may encounter TLS errors on `git push` — delegate all push/PR operations to the lead. - Each teammate runs in its own process context. - Worktree isolation is recommended for independent filesystem changes. ## Quick Reference ### Workflow Checklist - [ ] `CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1` set (the implicit team exists; no create step) - [ ] Task tools opted in when the lead runs on Opus 4.8+/Sonnet 5/Fable 5.x/Mythos 5 (`.claude/rules/agentic-permissions.md` § "Task-tool availability"); verify `TaskList` is callable before the first `TaskCreate` - [ ] `TaskCreate` for each work unit - [ ] Spawn teammates via the `Agent` tool with a `name` (no `team_name` needed) - [ ] `TaskUpdate` to assign tasks (or let teammates self-assign) - [ ] Receive messages automatically; respond via `SendMessage` - [ ] `SendMessage shutdown_request` to each teammate when done (no `TeamDelete`) ### Key Paths | Path | Contents | |------|----------| | `~/.claude/teams/` | Implicit-team state (members: name, agentId, agentType) | | `~/.claude/tasks/` | Shared task list state | Address teammates by the `name` you gave them at spawn — that is the reliable handle, independent of any on-disk layout. ### Common Mistakes | Mistake | Correct Approach | |---------|-----------------| | Using agentId as recipient | Use the `name` given at spawn | | Calling the removed `TeamCreate`/`TeamDelete` | The team is implicit (2.1.178); spawn with `Agent`, shut down with `shutdown_request` | | Passing `team_name` and expecting routing | It is accepted but ignored — there is one implicit team | | Sending broadcast for every update | Use `message` for single-recipient comms | | Polling for messages | Messages delivered automatically — just wait | | Sending JSON status messages | Use `TaskUpdate` for status, plain text for messages | | Sub-agent pushes to remote | Delegate push to lead orchestrator | ## Related Skills and Rules - `parallel-agent-dispatch` — worktree preflight, scope budgets, and the Return Contract every teammate must emit on exit. Team dispatches are a superset of plain parallel fan-out; follow both. - `.claude/rules/agent-development.md` — agent file structure, model selection, worktree isolation - `.claude/rules/agentic-permissions.md` — granular tool permission patterns - `.claude/rules/sandbox-guidance.md` — web sandbox constraints and push delegation