| name | team |
| description | Create an Agent Team from a predefined roster. Spawns multiple Claude Code sessions that coordinate via shared tasks and direct messaging. |
| context | fork |
| agent | team-lead |
| allowed-tools | Bash(git diff:*), Bash(git log:*), Bash(git show:*), Bash(git status), Bash(git branch:*), Glob, Grep, Read, Task, SendMessage |
| tags | ["meta"] |
| model | opus |
Agent Teams
Spawn a coordinated team of Claude Code sessions from a predefined roster. Each teammate is a full, independent session with its own context window. Teammates communicate directly with each other and coordinate through a shared task list.
Context
AVAILABLE ROSTERS:
- review-team (security-review + pragmatic-code-review + chaos-engineer)
- implementation-team (tdd-enforcer + refactoring-specialist + documentation-sync)
- investigation-team (3x systematic-debugger)
- design-team (system-design-reviewer + chaos-engineer + pragmatic-code-review)
- plan-review-team (system-design-reviewer + chaos-engineer + scope-analyzer)
GIT STATUS:
!`git status --short 2>/dev/null || echo "Not a git repo"`
CURRENT BRANCH:
!`git branch --show-current 2>/dev/null || echo "Unknown"`
Usage
/team review PR#142
/team implementation "add OAuth2 support to the auth module"
/team investigation "app exits after one message instead of staying connected"
/team design "we need to add multi-tenancy to the auth module"
/team plan-review .ai/sessions/2026-03-19_whatever/PLAN.md
Process
Step 1: Select Roster
Parse the user's command to identify the requested preset and task context.
Available presets:
review — security + code quality + chaos engineer (PR reviews, pre-merge audits)
implementation — TDD + refactoring + docs sync (feature work, large refactors)
investigation — 3x systematic debugger with competing hypotheses (bug hunting)
design — architect + adversary + code-quality reviewer (architecture debates before planning)
plan-review — architecture + risk + scope auditor (stress-test an existing PLAN.md)
If the preset is ambiguous or not specified, ask the user to choose.
Step 2: Load Roster
Read the roster file from teams/rosters/<preset>-team.yaml. Extract:
- Teammate definitions (name, agent, model, focus, plan_approval)
- Lead model and coordination strategy
- Team size
Present the team composition to the user:
- Who each teammate is and what they focus on
- The coordination strategy
- Expected token cost (each teammate ≈ 200k tokens; total ≈ team_size × 200k + lead)
Wait for explicit user approval before spawning.
Step 3: Create Team
Use Agent Teams primitives to:
- Create the team
- Create tasks based on the user's request, decomposed per the coordination strategy
- Spawn each teammate with their focus area and the agent definition from
agents/<agent>.md
Each teammate receives:
- Their
focus field as primary directive
- The task context from the user's request
- Access to the shared task list
Step 4: Coordinate
The team lead (this session) coordinates by:
- Monitoring teammate progress via task list
- Proactively pushing cross-teammate findings — when one teammate discovers something relevant to another, message them immediately
- Requiring plan approval from teammates where the roster specifies
plan_approval: true
- Reassigning work if a teammate gets stuck
Step 5: Synthesize
When all teammates finish:
- Collect results from each teammate
- Identify areas of agreement and disagreement
- Produce a unified report with per-teammate contributions
- Clean up the team
References