multi-agent-driven

Turn the current task into a coordinated multi-agent run — but only when fan-out actually pays. You stay the orchestrator: decompose, dispatch, integrate. Do not do the subtasks inline.

Language: mirror the user's language.

Gate: is fan-out justified?

Multi-agent costs are real: agents burn several times the tokens of a chat, multi-agent systems an order of magnitude more, and every hand-off pays a communication tax — a sub-agent's summary hides the implicit decisions behind its actions, and orchestrator instructions lose nuance on the way down.

1. Exhaust the single context first: clearer instructions, compressed context, a narrower tool set. Many "needs multi-agent" problems are single-agent context management done badly.
2. Fan out only on real signals: the current context is degrading (noise from failed explorations polluting decisions), the information space is wider than one agent can cover, or the tool set is large enough that selection accuracy drops and it splits along clear specialties.
3. When unsure, don't split. A wrong split (information loss plus inconsistent decisions) costs more than a slightly bloated single context. Start with the simplest approach that works; add agents only when evidence supports it.

Steps