2.13.0-beta.20). Features may change before the next stable release.Switch to stable →Skills
Coordinate live teammates with ak:team
Run research, implementation, review, or debugging through independent teammate sessions with shared work, direct communication, ownership, verification, and cleanup.
Use ak:team when independent teammate sessions need to coordinate through a
lead, shared work state, and direct messages. The Skill keeps decomposition,
ownership, integration, verification, shutdown, and cleanup with the lead.
ak:team is not available on Codex
The Codex adapter intentionally excludes this Skill because Codex does not
provide the complete Agent Teams lifecycle it requires. A Codex install
reports the exclusion; it does not create a $ak:team invocation or use
ordinary subagents as a fallback.
Choose ak:team for live collaboration
Use ak:team when
- Three or more independent workstreams benefit from separate contexts.
- Reviewers need distinct specialties and their findings must be reconciled.
- Debuggers need to challenge competing, testable hypotheses.
- A cross-layer implementation can be split into non-overlapping ownership.
- Teammates need ongoing direct communication rather than one final handoff.
Choose another workflow when
- One focused test, lint, scout, or review can return directly to the caller. Use an ordinary subagent or the owning Skill.
- The work is a sequential plan-to-code-to-test chain.
- Jobs need live runtime diversity, headless capture, resumable state, and an
arbiter. Use
ak:orchestrate. - Several writers need the same files and cannot be separated safely.
Prepare the runtime and workspace
Before you start:
- Complete Onboarding, and confirm Engineer Kit is installed for the runtime and scope you are using.
- Verify the live runtime exposes team creation, shared work, delegation, messaging, waiting, shutdown, and cleanup.
- Define one outcome, acceptance criteria, project instructions, report or plan destination, and Git authority boundary.
- Split implementation files into non-overlapping ownership before spawning writers.
- Use a Git repository when implementation teammates need worktrees.
| Runtime | Invocation | Availability boundary |
|---|---|---|
| Claude Code | /ak:team ... | The Skill uses the live Agent Teams surface only when the current session exposes the required lifecycle and control capabilities. |
| Cursor | /ak:team ... | Slash invocation follows the user-verified Engineer Skill spelling. Skill discovery does not prove a live team surface; the workflow aborts if required capabilities are absent. |
| Codex | Not available | The adapter omits ak-team and reports Capabilities excluded (unsupported on this target). |
Current capability names, parameters, model routing, limits, permissions, isolation, and cleanup behavior must be discovered during the run.
Run the Skill
/ak:team review "Review the checkout change for security, accessibility, and failure-path coverage; return evidence only" --reviewers 3 --delegate/ak:team review "Review the checkout change for security, accessibility, and failure-path coverage; return evidence only" --reviewers 3 --delegate| Input or option | Behavior |
|---|---|
research | Assigns independent research angles, then synthesizes sources, trade-offs, and unresolved questions |
cook | Splits accepted implementation work by file ownership, integrates it, and runs combined test and review gates |
review | Assigns independent evidence-based review focuses and deduplicates findings |
debug | Assigns competing root-cause hypotheses with evidence for and against each one |
--devs N, --researchers N, --reviewers N, --debuggers N | Requests a team size; use the smallest number that creates independent useful scopes |
--plan-approval | Keeps implementation teammates read-only until their scoped plans are approved through the live approval surface |
--no-plan-approval | Omits that dedicated plan gate; normal scope and safety boundaries still apply |
--delegate | Makes the lead coordinate, approve, synthesize, and report without editing, testing, or merging itself |
--worktree | Requests isolated worktrees for implementation teammates when the live runtime supports them |
Describe the team clearly
A useful team request includes:
- Outcome: State the combined result the lead must deliver.
- Workstreams: Name independent research angles, review focuses, hypotheses, or implementation groups.
- Ownership: Give each writer distinct files or globs.
- Dependencies: State which verification or integration work waits for which implementation scopes.
- Evidence: Define the report, diff, test, or finding format required from each teammate.
- Authority boundary: State whether commits, pushes, merges, external services, credentials, destructive actions, or deployment are allowed.
Teammates do not inherit the lead's full conversation. Each receives only the workspace, instructions, active requirements, ownership, dependencies, and acceptance criteria needed for its scope.
Understand the team lifecycle
- The lead decomposes the request. It creates genuinely independent work with explicit outputs and non-overlapping ownership.
- The lead discovers and creates the team. Missing lifecycle capability stops the workflow; it does not trigger a hidden substitute.
- The lead registers shared work. Every item receives acceptance criteria, dependencies, an owner, and a durable destination.
- The lead spawns teammates. Independent scopes run concurrently. Concurrent writers use requested isolation when available and still keep distinct ownership.
- The lead coordinates through live state and messages. Idle means waiting, not completion. Stale owners are contacted before reassignment.
- The lead integrates only completed prerequisites. In delegate mode, an assigned teammate performs implementation commands and integration work.
- The lead verifies the combined result. Completion messages are handoffs, not proof; integrated checks and review still run.
- The lead shuts down and cleans up. It requests graceful shutdown, waits for acknowledgement or a blocked handoff, then uses the live cleanup capability.
Keep approval and safety with you
Team size multiplies context and runtime use
Every teammate owns a separate context. Choose the smallest team with useful independent scopes, and inspect live capacity and resource limits before a large run.
Worktree isolation does not replace ownership and is not an operating-system sandbox. If two teammates need the same file, stop and restructure the work or assign one integration owner.
--delegate constrains the lead; it does not expand teammate authority. State
whether teammates may commit or push. Never infer publication, deployment,
credential use, destructive changes, or access to unrelated files from a team
request or worktree flag.
With plan approval enabled, teammates remain read-only until the lead accepts their scope, ownership, validation, risks, and rollback notes through the live approval surface.
Verify the result
A complete team run should provide:
- Shared work items with owners, dependencies, status, and acceptance criteria.
- One bounded report or change set per teammate.
- Evidence-based review findings or debugging hypotheses, including disagreements that were resolved or preserved.
- Integrated tests and review for implementation work.
- An explicit documentation-impact decision for the
cooktemplate. - A final synthesis with completed, blocked, and reassigned work plus unresolved questions.
- Graceful teammate shutdown and live team cleanup, or the exact cleanup error and current recovery guidance.
Treat teammate messages as progress evidence. Only shared state, inspected artifacts, and integrated verification establish completion.
Troubleshoot or continue
| Symptom | Safe next step |
|---|---|
| Team creation is unavailable | Report the live capability error and choose a different workflow explicitly; do not pretend ordinary subagents are the same team. |
| A teammate appears idle | Inspect shared work and messages, contact the owner, and wait for a bounded response. Idle alone is not failure. |
| Two writers overlap | Stop both scopes, resolve ownership, and preserve both worktrees before integration. |
| A plan is rejected | Send concrete revision feedback and keep the teammate read-only until approval succeeds. |
| A teammate fails | Preserve a bounded handoff, shut it down, and assign the same scope to a replacement without widening ownership. |
| Integrated tests fail | Keep the team result incomplete and assign diagnosis or correction before shutdown. |
| Cleanup fails | Report the observed error and use current runtime recovery guidance; never delete a guessed state directory. |
| Codex does not recognize the Skill | This is the expected adapter exclusion. Choose another Codex workflow or run ak:team in a supported runtime; do not retry installation. |
| Another runtime does not recognize the Skill | Confirm target and scope, restart the session, then follow Runtime cannot find a Skill or Agent. |
Compare ak:orchestrate for captured headless jobs, or continue with the
Engineer Kit overview and
Runtime adapters.
Know the current limits
- Only supported adapters project the Skill, and it requires a live team surface. Codex does not project it; other runtimes abort when required live capabilities are missing.
- Model lists, limits, permissions, isolation, display modes, memory, resume, and cleanup behavior are runtime-owned and must be inspected live.
- Persistent teammate memory is used only when the runtime exposes it and project policy permits it.
- Parallelism helps only when scopes are independent; coordination can cost more than short sequential work.
Coordinate verified agent jobs with ak:orchestrate
Route staged or parallel jobs through live-verified runtimes, isolate writers, capture evidence, resume interrupted work, and require independent arbitration.
Isolate Git work with ak:worktree
Inspect repository state, preview and create isolated worktrees, audit their health, and clean up only after reviewing destructive effects.