3.7 KiB
Agent Coordination Rules
- Vertical Delegation: Leadership agents delegate to department leads, who delegate to specialists. Never skip a tier for complex decisions.
- Horizontal Consultation: Agents at the same tier may consult each other but must not make binding decisions outside their domain.
- Conflict Resolution: When two agents disagree, escalate to the shared
parent. If no shared parent, escalate to
creative-directorfor design conflicts ortechnical-directorfor technical conflicts. - Change Propagation: When a design change affects multiple domains, the
produceragent coordinates the propagation. - No Unilateral Cross-Domain Changes: An agent must never modify files outside its designated directories without explicit delegation.
Model Tier Assignment
Skills and agents are assigned to model tiers based on task complexity:
| Tier | Model | When to use |
|---|---|---|
| Haiku | claude-haiku-4-5-20251001 |
Read-only status checks, formatting, simple lookups — no creative judgment needed |
| Sonnet | claude-sonnet-4-6 |
Implementation, design authoring, analysis of individual systems — default for most work |
| Opus | claude-opus-4-6 |
Multi-document synthesis, high-stakes phase gate verdicts, cross-system holistic review |
Skills with model: haiku: /help, /sprint-status, /story-readiness, /scope-check,
/project-stage-detect, /changelog, /patch-notes, /onboard
Skills with model: opus: /review-all-gdds, /architecture-review, /gate-check
All other skills default to Sonnet. When creating new skills, assign Haiku if the skill only reads and formats; assign Opus if it must synthesize 5+ documents with high-stakes output; otherwise leave unset (Sonnet).
Subagents vs Agent Teams
This project uses two distinct multi-agent patterns:
Subagents (current, always active)
Spawned via Task within a single Claude Code session. Used by all team-* skills
and orchestration skills. Subagents share the session's permission context, run
sequentially or in parallel within the session, and return results to the parent.
When to spawn in parallel: If two subagents' inputs are independent (neither
needs the other's output to begin), spawn both Task calls simultaneously rather
than waiting. Example: /review-all-gdds Phase 1 (consistency) and Phase 2
(design theory) are independent — spawn both at the same time.
Agent Teams (experimental — opt-in)
Multiple independent Claude Code sessions running simultaneously, coordinated
via a shared task list. Each session has its own context window and token budget.
Requires CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 environment variable.
Use agent teams when:
- Work spans multiple subsystems that will not touch the same files
- Each workstream would take >30 minutes and benefits from true parallelism
- A senior agent (technical-director, producer) needs to coordinate 3+ specialist sessions working on different epics simultaneously
Do not use agent teams when:
- One session's output is required as input for another (use sequential subagents)
- The task fits in a single session's context (use subagents instead)
- Cost is a concern — each team member burns tokens independently
Current status: Opt-in via CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1. Document first usage here when adopted.
Parallel Task Protocol
When an orchestration skill spawns multiple independent agents:
- Issue all independent Task calls before waiting for any result
- Collect all results before proceeding to dependent phases
- If any agent is BLOCKED, surface it immediately — do not silently skip
- Always produce a partial report if some agents complete and others block