away-team

name	away-team
description	Use for complex implementation tasks requiring coordinated multi-agent execution. Trigger: away team, team up, multi-agent implementation, or tasks too large for a single agent.
tags	["multi-agent","team","workflow","orchestration"]

/away-team — Standard Multi-Agent Team Pattern

Spawn a coordinated team for complex implementation tasks. The team lead coordinates, engineers implement, and reviewers validate.

Arguments

• /away-team "<task description>" — Standard team (lead + 2-3 agents)
• /away-team --small "<task>" — Minimal team (lead + 1 agent)
• /away-team --full "<task>" — Full team (lead + 3 agents + reviewer)

Team Composition

Standard Team (default)

Role	Model	Agent Type	Purpose
Team Lead	Opus	general-purpose	Coordinates, never implements directly
Engineer	Sonnet	sonnet-worker	Implements code changes
Security Officer	Sonnet	sonnet-reviewer	Reviews security implications
Science Officer	Haiku	haiku-explorer	Researches patterns and prior art

Small Team (--small)

Role	Model	Agent Type	Purpose
Team Lead	Opus	general-purpose	Coordinates + reviews
Engineer	Sonnet	sonnet-worker	Implements all changes

Full Team (--full)

Standard team plus:

Role	Model	Agent Type	Purpose
Quality Officer	Sonnet	sonnet-reviewer	Runs adversarial review on results

Workflow

flowchart TD
    Start([away-team]) --> Parse{Parse team size}
    Parse -- "--small" --> SmallPlan["1: Mission Planning\nLead + 1 Engineer only"]
    Parse -- "default / --full" --> Plan["1: Mission Planning\nAnalyze task, create subtasks\nvia TaskCreate, identify disjoint file sets"]
    SmallPlan --> SmallImpl["2-4: Lead assigns + Engineer implements"]
    SmallImpl --> SmallReview["5: Lead reviews\n(no Security/Quality Officer)"]
    SmallReview --> Report
    Plan --> Intel["2: Intelligence Gathering\nScience Officer searches codebase\n+ queries knowledge graph"]
    Intel --> Assign["3: Task Assignment\nTaskUpdate with file scopes\n+ addBlockedBy dependencies"]
    Assign --> Impl{{"4: Implementation\nEngineers execute assigned tasks\n(parallel, disjoint file sets)"}}
    Impl --> SecReview["5a: Security Officer reviews\neach implementation + Devil's Advocate"]
    SecReview --> FullTeam{--full flag?}
    FullTeam -- Yes --> QualReview["5b: Quality Officer runs\nadversarial debate on final result"]
    FullTeam -- No --> Report
    QualReview --> Report["6: Team Report"]
    Report --> Cleanup["7: Shutdown all agents\nvia shutdown_request"]
    Cleanup --> Done([Done])

Key Decisions

Mission Planning: Team Lead analyzes task, breaks into subtasks with TaskCreate. File sets MUST be disjoint — no file assigned to multiple agents. Team Lead coordinates but never implements directly.

Intelligence Gathering: Science Officer (haiku-explorer) searches codebase for relevant patterns, checks knowledge graph for prior decisions, reports findings to Team Lead before task assignment.

Task Assignment: Each agent gets a clear scope with specific file paths. Include full file contents in agent prompts (agents have no shared context). Set dependencies via addBlockedBy so reviews block on implementation.

Review: Security Officer reviews each completed task, runs Devil's Advocate on design decisions, flags security concerns with severity ratings. Quality Officer (full team only) runs adversarial debate producing confidence-classified findings.

Team Report format:

## Away Team Report — {timestamp}

**Mission**: {task description}
**Team**: {composition}
**Duration**: {time}

### Tasks Completed
| Task | Agent | Status | Files Modified |
|------|-------|--------|---------------|
| {task} | {agent} | {done/partial} | {files} |

### Review Findings
| Finding | Severity | Confidence | Reviewer |
|---------|----------|------------|---------|
| {finding} | {sev} | {HIGH/MED/LOW} | {who} |

### Decisions Made
- {decision 1 with rationale}

### Mission Status
{COMPLETE / PARTIAL — with explanation}

Critical Rules

1. Disjoint file sets: NEVER assign the same file to multiple agents
2. Full context in prompts: Agents start fresh — include all needed file contents
3. Verify by file existence: Don't rely solely on task completion status
4. Set Bash timeout: Use 600000ms for long operations
5. Hull integrity monitoring: If any agent hits Red (33%+ context), trigger Relief on Station

Troubleshooting

Symptom	Cause	Fix
Engineer agent can't run Bash commands	sonnet-worker lacks Bash tool permissions	Return Bash-requiring tasks to Team Lead (main agent) for execution
Two agents edit the same file	File sets not disjoint in task assignment	Re-assign: each file must belong to exactly one agent
Agent has no context about the codebase	Agent starts fresh; didn't receive file contents	Include full file contents in agent prompt — agents cannot share context
Team Lead tries to implement directly	Role confusion in prompt	Team Lead coordinates only; redirect implementation to Engineer agents
Agent hits Red hull integrity (33%+) mid-task	Large files or complex task consuming context	Trigger /relief-on-station for that agent; replace with fresh agent

Notes

• Team uses TeamCreate for coordination infrastructure
• JAVA_HOME on MINGW: export JAVA_HOME="/c/Program Files/Eclipse Adoptium/jdk-17.0.17.10-hotspot"
• sonnet-worker agents may have limited Bash capabilities — plan accordingly
• Token cost: ~20,000-40,000 tokens total (team lead + 3 agents + reviews)
• Always-on overhead: 0 (loaded on demand)