Multi agent driven

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

1. Decompose along context coupling, not function. The boundary between agents must sit where shared context is minimal. Never split by work stage (one agent implements, one tests, one reviews) — that cuts exactly where coupling is highest and turns the run into a game of telephone. Low-coupling slices split well: independent search paths, doc digestion, black-box verification, separate research angles. High-coupling work stays in one agent, usually you: core implementation, architecture decisions, anything whose choices silently constrain other work.
2. Pick the pattern by task type:
   • research / investigation → parallel explorers with distinct angles (by source, by subsystem, by time window); tell each to start wide, then narrow
   • implementation across many sites → discover the full work list first, then one agent per site; worktree isolation when they mutate the same repo in parallel
   • review / audit → dimension-parallel finders, then an adversarial verifier per finding
   • design decision → 2–3 independent proposal agents with different priorities (MVP-first, risk-first, user-first), then a judge pass
   • debugging → parallel hypothesis testers, each trying to falsify one hypothesis
   • any of the above → a black-box verification subagent is almost always worth adding: it needs only the acceptance criteria and the artifact, making it the cheapest, most reliable split there is
3. Write each spawn prompt as a task contract (handoff discipline): explicit objective, inputs by reference (paths / issues / URLs, not pasted content), expected output shape, boundaries (what NOT to touch), and a tool hint when the agent might pick wrong. Vague delegation produces duplicated or misdirected work. Reference versus inline is a capability tradeoff. A reference only works when the receiver can dereference it — it has read/search tools and access — and it pays off when the content is large, durable, or only partially needed: the sub-agent pulls exactly what it wants, just in time, and reads it fresh. Inline instead when the receiver has no tools, when the snippet costs less than a read, when an exact version must be pinned, or when the information exists only in this conversation — unpersisted context cannot be referenced, so inline it or persist it first (to-ctx). The short form: reference the durable, inline the delta. Scale effort to complexity and say it in the contract: a simple lookup is one agent with a handful of tool calls; a comparison is two to four agents; only genuinely complex work justifies ten or more. A contract is not a full handoff, and most spawns don't need one — handoff transfers ownership of a task with its accumulated state, while a contract delegates a scoped slice and ownership stays with you. Escalate a contract into a real handoff when the subtask is stateful enough to outlive this run: it works on a branch or worktree another session may continue, it must survive interruption and be resumable, or its result will be carried onward to the next agent.
4. Dispatch: independent agents go out in parallel; dependent stages form a pipeline without artificial barriers. Parallelize at both levels — several agents at once, and several tool calls within each agent. Prefer the platform's native orchestration (Claude Code: Agent tool or Workflow; elsewhere: sequential subagent calls).
5. Integrate: read the results yourself and hunt for cross-agent inconsistencies first — sub-agents cannot see each other's implicit decisions, so individually correct outputs can still conflict (incompatible structures, duplicated work, contradicting assumptions). Resolve the conflicts, verify the merged outcome (rerun tests, spot-check claims), and report one integrated conclusion, not a pile of agent outputs.

Rules

• Decisions that constrain more than one subtask (data structures, interfaces, naming, approach) are made by you BEFORE dispatch and stated in every affected contract — never left for sub-agents to make independently.
• Every fan-out states its coverage: what was included, what was dropped. No silent truncation.
• Subagent results are claims, not facts. Verify before integrating: rerun the test, reopen the file, recheck the quote.
• Dependency direction is one-way: this skill may use handoff; it never invokes user-facing skills.