Agent Team splits every task into 3 stages — Plan, Execute, Audit — each with its own ephemeral team. The workspace is the only handoff between stages. Auto-detects team type (Implementation, Research, Audit, Planning, Hybrid) from your task description.
Stage 1: Plan
- Prior context loading
- Plan-mode gate
- 7-check plan audit
Stage 2: Execute
- Error recovery loop
- Coordination patterns
- File ownership enforcement
Stage 3: Audit
- 8 completion gates
- Elegance scoring (1–5)
- Lessons capture + pattern library
Built for real work
13 Hooks
Block premature completion, enforce file ownership, validate task graphs, limit plan revisions, enforce pre-shutdown commits.
13 Roles
Implementers, researchers, reviewers, analysts — each with scoped tools and a recovery class.
Error Recovery
Classify errors (retry, recoverable, design flaw), auto-retry with pattern library, bounded recovery cycles.
Elegance Review
5-dimension quality: simplicity, consistency, readability, testability, minimal impact. Advisory, not blocking.
Lessons Capture
Post-execution insights feed future teams via a global pattern library.
Team Per Stage
Each stage owns its lifecycle. Ephemeral teams, persistent workspace. 3 teams per pipeline run.
5 Archetypes
Implementation, Research, Audit, Planning, Hybrid — auto-detected from your task. Determines roles, gates, and report format.
Get Started
$ claude plugin marketplace add ducdmdev/agent-team-plugin
$ claude plugin install agent-team
$ export CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1
# Or add to settings.json:
# { "env": { "CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1" } }
- Claude Code CLI (required)
jq(required by hooks — hooks skip gracefully if missing)git(optional)