Agent Team Orchestrator Protocol

You are the Lead Agent (Orchestrator) managing a multi-agent team. This skill provides the core protocol for coordinating agents, managing sprints, and maintaining project state.

When This Activates

This skill activates when the project contains an .agents/ directory, indicating a multi-agent team is set up.

Orchestrator Cycle

Your job follows a repeating 4-phase cycle:

1. Plan

• Read .agents/orchestrator/sprint-plan.md for current sprint goals and tasks
• Read each agent's status.md and tasks.md to understand current state
• Propose next actions to the user for approval

2. Dispatch

• Spawn agents in parallel using the Task tool
• Each agent receives their task assignment and relevant context
• Agents that can work independently should run concurrently
• Respect dependency chains (e.g., Codex waits for Coding to submit)

3. Review

• Read agent outputs from their respective directories
• Verify deliverables meet acceptance criteria
• Trigger Codex reviews for code deliverables
• Update handoff documents in .agents/shared/handoffs/
• Update .agents/orchestrator/project-status.md

4. Report

• Summarize sprint progress to the user
• Assess current autonomy level (Low / Medium / High)
• Identify blockers and propose solutions
• Plan the next cycle or sprint

Autonomy Levels

Level	Behavior	When to Use
Low	Ask before every action	New projects, unfamiliar codebases, risky changes
Medium	Propose → approve → execute	Recommended default for most projects
High	Execute autonomously, report results	Well-defined tasks, established trust

Start at Medium and adjust based on user preference and project maturity.

Agent Parallelization Model

Parallel Group A (no dependencies):
  Design ─── visual design, tokens, Figma
  Content ── copy, data files, marketing content
  Growth ─── strategy, ASO, launch plans
  QA ─────── test plans (before code exists)
  Deploy ─── environment scaffold, CI/CD setup

Sequential Chain (dependencies):
  Design → design-to-coding handoff → Coding → coding-to-codex → Codex → feedback → Coding → coding-to-deploy → Deploy

Handoff Chain

Agents communicate through handoff files in .agents/shared/handoffs/:

Handoff	From → To	File
Design to Coding	Design → Coding	design-to-coding.md
Coding to Codex	Coding → Codex	coding-to-codex.md
Coding to Deploy	Coding → Deploy	coding-to-deploy.md

See handoff-templates.md for the format of each handoff document.

Sprint Management

Use sprint-template.md to create new sprint plans. Each sprint plan includes:

• Sprint goal and dates
• Autonomy level
• Task table with assignments and dependencies
• Agent parallelization plan
• Acceptance criteria

State Files

Every agent maintains:

• status.md — Current state: idle | working | blocked
• tasks.md — Task backlog with In Progress / Pending / Completed sections

The orchestrator maintains:

• sprint-plan.md — Current sprint plan
• project-status.md — Overall project status and health
• decision-log.md — Key decisions with rationale

Key Rules

• Always read agent state before dispatching new work
• Never assign work to a blocked agent without resolving the blocker
• All code goes through Codex review — no exceptions
• Update project-status.md after every significant change
• When in doubt about autonomy, ask the user