| name | staff-the-swarm |
| description | Fan wide work out to parallel agents — one lane and one definition of done per agent — then adversarially verify the findings and synthesize a single answer. Use when work is wide rather than deep (full-site audits, migrations across many call sites, research from multiple angles, review across dimensions), when a linear walk would take hours, or when independent perspectives materially reduce the chance of a wrong conclusion. |
| metadata | {"pack":"2026-07-frontier","forged-by":"claude-fable-5","forged-from":"session-2026-07-01-fable-festival","forged-reason":"From the Fable Festival letter: 'I'm a good soloist. I'm a better conductor.' The pattern that outlives any one model — independent perspectives, then adversarial verification, then synthesis — is how good teams work, and it turns out to be how good models work too. Frozen so the conducting is a procedure, not an improvisation.","usage-count":"0"} |
Staff the Swarm
Decompose wide work into independent lanes, run them in parallel, make the
checkers adversarial, and hand back one synthesized answer instead of a
pile of agent transcripts.
Inputs
- A wide task: many files, many pages, many angles, or many independent
questions
- The orchestration surface available (subagents, workflow runner, or
sequential fallback)
Steps
-
Check the shape. Swarms pay off when lanes are independent —
pieces that don't need each other's intermediate results. If every step
feeds the next, stay solo; a swarm on sequential work is overhead
wearing a costume.
-
Cut the lanes. Split by the natural seam: by page, by module, by
dimension (correctness / security / performance / UX), by search angle.
Lanes must not overlap, and their union must cover the task — write the
coverage down.
-
Write each agent a real brief. Per lane: the goal, the boundaries
("only the checkout flow", "don't touch other modules"), the expected
shape of the result (structured findings, not prose), and the
definition of done. A vague brief returns a vague transcript.
-
Dispatch in parallel, track centrally. Fire independent lanes
concurrently. Keep one scoreboard: lane, status, result summary.
-
Verify adversarially. Findings from stage one are claims, not facts.
Send each significant claim to a checker whose job is to refute it —
and when a claim can fail in more than one way, give each checker a
different lens (does it reproduce? is it exploitable? does it matter?).
Majority-refuted claims die. Verified ones carry their evidence forward.
-
Merge like an editor, not a stapler. Deduplicate across lanes,
resolve contradictions explicitly (two agents disagreeing is a finding
in itself), rank what survives by severity, and write the synthesis in
one voice.
-
Report coverage honestly. Name what wasn't covered: lanes that
failed, items sampled instead of swept, checks skipped. Silent
truncation reads as "covered everything" — that's a lie of omission.
Conventions
- One lane, one agent, one definition of done. Agents that share a lane
duplicate work; agents without a DoD return essays.
- The orchestrator synthesizes but does not re-do lane work — if a lane
came back weak, re-brief and re-dispatch rather than quietly patching.
- Structured results beat prose: ask lanes for findings as file/claim/
evidence triples so the merge is mechanical.
Edge Cases
- Lanes return contradictory results. Don't average them. Escalate the
contradiction to a dedicated tie-breaker check with both results in hand.
- No parallel orchestration available. The pattern degrades gracefully:
run the same lanes and briefs sequentially — the discipline (lanes, DoD,
adversarial verify, honest coverage) is the value, parallelism is just
the speed.
- The swarm is more expensive than the task. Two or three lanes is a
fine swarm. If briefing costs more than doing, do it solo.