multi-agent-orchestrationlisted

# Multi-agent orchestration Source: [How we built our multi-agent research system](https://www.anthropic.com/engineering/multi-agent-research-system). The pattern lives in Anthropic's cookbook as notebooks and prompts. This packages the parallel-research recipe as a skill, with the judgment of when to use it. The pack also ships a `/research` command and a `research-worker` subagent that run this flow. A lead agent plans a question, spawns several subagents to chase parts of it in parallel, and synthesizes what they return. Each subagent has its own context window, so the system explores far more ground than one agent could hold at once. That power has a price, so the first decision is whether to use it at all. ## Use it for breadth, not for coupling Multi-agent fits work that splits into independent pieces explored at the same time: surveying a literature, gathering evidence from many sources, mapping a large unknown space. It fits poorly when the pieces depend on each other. Do not reach for it on a coding task with shared state, or anything that needs tight coordination between the parts, because the subagents cannot see each other's context and will step on the shared thing. A multi-agent run also costs on the order of 15 times the tokens of a single chat. Spend that only when the breadth is worth it. ## Delegate so workers do not collide A vague subagent prompt produces overlap and gaps. Give each worker four things: - a specific objective, narrow enough that two