Orchestrating Subagents

Orchestrating Subagents

Swarm-mode work in this repo delegates heavily (explorer → reviewer → critic). This skill defines HOW to delegate so validation gates stay strong while breadth stays fast and cheap. It complements — never replaces — the gates in .claude/session/swarm-mode.md, qa-sweep, and swarm-implement.

Role → tier mapping

Hard rule: economize on explorers, never on reviewers or critics. If the harness exposes model or effort overrides for subagents, tier explorers down; do not tier the reviewer or critic below the session model or below high effort. If no override is available, tier by agent type (Explore is lightweight: read-only tools, skips CLAUDE.md) and by prompt scope.

Fan-out discipline

• Launch parallel agents only for disjoint scopes. Before launching, write one line per agent stating its scope; if two overlap, merge them.
• 2–4 explorers per wave is a heuristic only when the active workflow does not define its own fan-out contract. A workflow-specific contract overrides this heuristic: PR review, for example, requires all six base dimensions and all eleven risk families to be covered on every PR — its capability profiles and depth tiers (see swarm-pr-review) govern how many lanes carry that coverage, and under the plugin's mechanical controller the tier is computed from the bound diff and the matching lane floors are enforced outright. No time, cost, repository-size, or simplicity rationale may reduce the coverage a workflow mandates. Outside fixed-fan-out workflows, more agents than distinct scopes adds token cost and synthesis burden without adding recall.
• Launch independent agents in a single message so they run concurrently.
• Do not re-run a search an agent is already doing; wait for its report.
• Scale waves, not width: if the first wave surfaces new territory, launch a second targeted wave rather than one giant speculative first wave.

Subagent prompt contract

Every delegation prompt must state:

1. Scope — exact directories, files, or question. Name what is out of scope.
2. Deliverable — the structure of the report (per-item findings, then a ranked summary). The agent's final message is the only thing you receive.
3. Evidence bar — exact file:line references; no invented paths; verify a path exists before citing it.
4. Status labels — explorers return CANDIDATE findings only; reviewers classify CONFIRMED / DISPROVED / UNVERIFIED / PRE_EXISTING; reviewer and critic verdicts are APPROVE / NEEDS_REVISION / BLOCKED.
5. Output bound — compact structured returns; no full-file dumps.

For reviewers and critics additionally:

• Give the claims and locations, not the author's justification — the reviewer must re-derive, not confirm, the reasoning.
• State the adversarial default explicitly: "default to DISPROVED/UNVERIFIED unless the code evidence supports the finding."
• A reviewer or critic must be a fresh agent, never a continued conversation with the agent whose work it judges.

Independence and staleness

• Reviewer and critic review the latest diff, not a description of it. Give them the branch state and the validation evidence, and require them to run git diff/git log themselves.
• Any edit after an approval invalidates it. Record what was approved (e.g. git rev-parse HEAD, git diff --stat) so staleness is checkable — see the durable-session-state skill.

Nesting limitation

Whether a subagent can spawn further subagents depends on the harness and agent type — check whether a subagent tool (Agent or Task) is actually available in your context before assuming either way. If a skill mandating fresh-subagent review (qa-sweep, swarm-implement) executes in a context without a subagent tool:

• perform the same review checklist yourself as a clearly labeled fallback self-review, and
• disclose in your report that independent review was unavailable in this context, so the orchestrator can re-run the gate with a real fresh agent. Never silently present self-review as independent review.

Main-context hygiene

• Push reading-heavy work into subagents; keep the main thread for scoping, synthesis, and decisions.
• Do not paste subagent transcripts or large tool outputs back into the main thread; carry forward only validated findings and verdicts.
• When a subagent report arrives, extract the load-bearing facts into your durable task artifacts (see durable-session-state) before moving on.

When not to delegate

Answer directly, without a subagent, when the task is a single-fact lookup you can resolve with one or two targeted Grep/Read calls, or when you already know the file and symbol. Delegation overhead should buy breadth, isolation, or independence — if it buys none of those, skip it.