blackms/aistack/orchestrate

aistack: Orchestrate a multi-agent run

aistack turns Claude Code into a multi-agent orchestrator. Instead of one generalist doing everything, you decompose work and route each part to a specialized agent backed by the bundled aistack MCP server.

When to use

• A task spans distinct skills (design + implement + test + review).
• You want a dedicated critic (adversarial / reviewer) separate from the implementer.
• The user explicitly asks to orchestrate, spawn agents, or run a workflow.

For a small, single-step change, do not orchestrate — just do it directly.

Available agent types

How to orchestrate

1. Decompose: break the task into 1–4 sub-tasks, each mapped to one agent type above.
2. Discover tools: the aistack MCP server exposes agent_types, agent_spawn, agent_list, agent_status, agent_stop, task_create, task_assign. List them with npx @blackms/aistack mcp tools if unsure.
3. Spawn: call agent_spawn with the chosen type and a focused prompt (include acceptance criteria and the relevant file paths). Spawn background where possible so you can fan out.
4. Coordinate: track progress with agent_list / agent_status. Pass artifacts between agents via shared memory (see the aistack-memory skill) rather than re-deriving them.
5. Converge: when sub-tasks finish, have a reviewer or adversarial agent validate the combined result before reporting back.

Tips

• Keep each agent's prompt narrow — one responsibility. Broad prompts dilute the benefit of specialization.
• Prefer opus-backed agents (architect, adversarial, security-auditor, coordinator) for judgment-heavy steps; sonnet for execution.
• Export agents as native Claude Code subagents with npx @blackms/aistack export-agents if you want to @-mention them directly.

Reference: https://github.com/blackms/aistack