fan-out

name

fan-out

description

Parallelize a large task across many subagents — when the work is independent per-unit (verify / research / generate, the same operation across N files or items) it is a fan-out shape, and parallel fresh-eyes-per-unit beats one tiring serial operator. Use to triage whether a long task should be fanned out, and to run the fan-out safely under an active presiding watch (e.g., a 62-file audit fanned to 6 agents finished in ~15 min and was more thorough than the serial walk that preceded it).

Fan-Out

The triage (do this BEFORE starting any task with >~20 independent units)

Ask one question: what is the SHAPE of this work?

• Fan-out shape — independent, per-unit, the same operation across many units (verify N files against sources, research N videos, generate N stubs). No unit needs another unit's result. → Fan out. Parallel speedup and a higher quality ceiling: fresh eyes per unit beat one operator who accumulates blind spots ("I already know the pattern, I'll stop re-checking").
• Single-pass shape — centralizable (one known pattern, deterministic to locate — e.g. "remove all the X notes"), or sequentially dependent (each step needs the last). → One careful pass. Fan-out here just adds coordination, consistency risk, and review cost without improving the output.

The classic miss: a verification walk across hundreds of files run serially for days because "just do the next file" momentum never triggered the triage. It was a textbook fan-out shape. Surface the triage result before starting — a quick "this is ~N independent units; I'd fan it out / I'd do it single-pass because ___."

Why fan-out has a higher ceiling on verification (the refined principle)

"The full-context shepherd is the ceiling" is only half true:

• Shepherd-for-integration — one full-context mind is the ceiling for cross-cutting issues (a race spanning files, a thesis contradicting another document). Fan-out can't see those.
• Fan-out-for-independent-verification — for per-unit checks, parallel fresh eyes BEAT one accumulating operator. They are complementary, not rivals.

Serial-probe, then parallel-scale. A short serial pass finds the pattern (e.g. "the dominant defect is quotes drawn from the wrong edition of the source; also check the citations inside entries"). Then front-load that knowledge into the fan-out spec so the parallel agents inherit the shepherd's accumulated context without the blind spot.

The recipe

1. Enumerate + batch. List the units; split into N batches (aim ~10-14 units/agent; balance by weight, not just count).
2. Write ONE shared spec. Self-contained — the agents start cold:
   • the why (one paragraph of context),
   • the exact method + tools (concrete commands / tool names),
   • the rules (read-before-quoting; flag-don't-resolve for sensitive or judgment-call items; conservative — correct only on a clear source mismatch, else FLAG),
   • the front-loaded pattern from the serial probe,
   • the output format (a compact structured per-unit report: corrections <ref>: "stale" → "genuine" (verified via X), flags, clean items, tally). Give each agent the shared spec + its batch slice. Tell them: edit + report, no git/commits (the presider commits after review).
3. Stage it — wave 1 to validate, then scale. Spawn a SMALL first wave (~2 agents) on varied batches. Review their work before spawning the rest. If the spec produced clean, correct output, spawn the remaining waves; if not, fix the spec cheaply (git reverts the small wave). This is the D&C 121 "reprove, then show increase of trust" applied to your own delegation.
4. Spawn in parallel. Multiple Agent calls in one message run concurrently (or run_in_background: true). Use a strong model for careful verification.

The presiding watch (non-negotiable)

A fan-out is a delegation. Your delegation covenant or standing instructions, if you keep one, govern it — watch what you order (Abraham 4:18 — watch until it obeys the INTENT, not until the job returns):

• Review every agent's report.
• Spot-verify the high-stakes catches against the source yourself — don't accept a confabulation-fix or a citation-correction on the agent's word; re-grep / re-query it. (In the worked example below, six spot-checks spread across all agents confirmed reliability before ~75 edits were accepted.)
• Check the diff is bounded (proportionate to the reported corrections; no wholesale rewrites).
• Consolidate: merge reports into the audit record (e.g. findings.md); fix any cross-cutting issues the units surfaced (a fan-out can reveal an error that reached a parent artifact — chase those down).

Fan-out without the watch is offloading; fan-out with the watch is multiplication. The watch is the real bottleneck, which is why the output format must be reviewable.

Cost

Spawning is the expensive path — each agent re-derives context cold. Justified when (parallel speedup + fresh-eyes-per-unit value) > (coordination + review + cold-start cost). True for independent verification across many units; false for centralizable single-pattern cleanup.

Worked example

A 62-file source audit, ~9-18 min in parallel. 6 subagents, staged 2-then-4. The shared spec front-loaded the serial probe's pattern (a wrong-edition defect class + "check the citations inside entries"). ~75 corrections; the presider reviewed all reports, spot-verified 6 high-stakes catches, and checked the diff before committing. It caught a defect class the serial walk under-checked (fabricated citations inside quoted entries) and one error that had reached a parent artifact the serial pass marked CLEAN — the strongest evidence for the higher ceiling.