Skip to content

feat: Agent Teams orchestration — /team skill + teammate-aware preamble#105

Open
HMAKT99 wants to merge 2 commits intogarrytan:mainfrom
HMAKT99:arun/agent-teams
Open

feat: Agent Teams orchestration — /team skill + teammate-aware preamble#105
HMAKT99 wants to merge 2 commits intogarrytan:mainfrom
HMAKT99:arun/agent-teams

Conversation

@HMAKT99
Copy link

@HMAKT99 HMAKT99 commented Mar 16, 2026

Summary

Makes every gstack skill work as a Claude Code Agent Team teammate. Skills auto-detect
if they're running as a teammate and switch behavior: messaging findings to other
teammates instead of outputting to the conversation.

Changes

  • Preamble: _IS_TEAMMATE detection added to all skills via {{PREAMBLE}}
  • /team skill: 7 pre-built team configurations (ship, review, launch, incident, diligence, audit, custom)
  • team/TEAMS.md: Coordination reference with dependency graph, message formats, shared state
  • CLAUDE.md: Agent Teams documentation section

How it works

/team diligence → spawns 5 teammates:
  vc (moat analysis) ──────────┐
  cfo (cost modeling) ─────────┤
  cso (security audit) → risk (incorporates findings) → board (synthesizes all)

Each teammate reads their SKILL.md, follows the methodology, and messages findings
to other teammates. The lead synthesizes outputs into a unified report.

Opt-in required

Requires CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 in user's settings.json.
Not committed to the repo — users enable it themselves.

Test plan

  • All existing tests pass with preamble changes
  • All regenerated SKILL.md files are FRESH
  • /team skill has valid frontmatter and preamble

Makes every gstack skill work as a Claude Code Agent Team teammate:

1. Preamble: _IS_TEAMMATE detection, communication protocol, task claiming
2. /team skill: 7 pre-built team configurations (ship, review, launch, incident, diligence, audit, custom)
3. team/TEAMS.md: coordination reference with dependency graph and message formats
4. CLAUDE.md: Agent Teams documentation section

Requires: CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 (user opts in)
Tier 3 evals (~$0.08/run) using Claude Sonnet as judge:
- /team SKILL.md orchestration quality (spawn prompts, team configs)
- TEAMS.md coordination reference quality (message formats, state locations)
- Preamble teammate awareness actionability (mode switching)
- Spawn prompt context sufficiency (enough info for autonomous teammates)

Run: EVALS=1 bun test test/agent-teams-llm-eval.test.ts
Requires: ANTHROPIC_API_KEY
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant