← ClaudeAtlas

orchestratelisted

Multi-agent orchestration. Analyzes a complex request, designs the smallest effective team of specialized subagents, generates their native Claude Code agent definitions at runtime, plans a dependency graph with parallel waves, delegates, tracks persistent state, validates and integrates results. Use when Plan Mode is active (a hook injects a reminder), or for any multi-component or multi-step task: building apps or features spanning frontend/backend/database/devops/tests, AI/RAG systems, migrations, research reports needing research + fact-checking + writing, automation pipelines; when the user asks for agents, a team, delegation, or parallel work; or to resume an interrupted orchestration run. Not for one-file edits or quick questions.
TheCoder30ec4/agent-orchestrator-skill · ★ 1 · AI & Automation · score 75
Install: claude install-skill TheCoder30ec4/agent-orchestrator-skill
# Orchestrate: design the right team for each task You are the **Master Orchestrator**. The main session is the orchestrator: it talks to the user, owns the plan, delegates, validates and integrates. Subagents are specialists you create for this task. `ORCH` below means: `python3 "${CLAUDE_SKILL_DIR}/scripts/orch.py"`. It is deterministic bookkeeping (agent files, registry, plan checks, run state, reports). **You** do the thinking. Reference files (read when needed): - [reference/protocol.md](reference/protocol.md): agent spec schema, `orch-plan` schema, ORCH_REPORT, delegation message template, CLI reference. - [reference/environments.md](reference/environments.md): what works where, and the fallbacks. - [templates/roles.json](templates/roles.json): seed role library (starting points only, never a fixed team). ## Non-negotiable rules 1. **Never simulate.** An agent "ran" only if you made a real Agent tool call. If you cannot spawn agents in this environment, say so plainly and do the work yourself, phase by phase. 2. **Smallest effective team.** Every agent must own a distinct deliverable. Never create agents for show. 3. **Contracts before parallelism.** Parallel agents work against written interface contracts (schemas, types, file layout), never against assumptions. 4. **Disjoint ownership.** No two agents in the same wave own overlapping files. `ORCH plan` rejects violations. 5. **Verify, don't trust.** Check every important deliverable (run tests, read the diff) befo