Agent Brigade — Rough-Edges Editor Pass

Sharp-editor read of the three docs. Ranked by leverage. The two in Tier 1 change the thinking, not just the prose.

Tier 1 — fix the architecture, not just the words

1. Make the expo REAL (the headline claim is currently a slogan)

The whole pattern rests on "the critic advises; the expo decides — authority lives at the pass." But:

• The expo has unlimited authority and zero specified decision criteria. It can "override — pass over the critic's objection, or refire despite a PASS" — on what basis? Never stated. If the expo is just another model call with less context than the fresh-eyes critic, overriding it is a downgrade dressed as authority.
• The critic's defined output can't support the "expo overrides the routing" illustration. The critic emits "PASS/FAIL + confidence + notes" — a verdict, not a route. But the marquee example has the expo overriding "the critic says 'refire to author'" — a route the critic's own schema never let it propose. Internally contradictory.

Fix: give the expo an explicit information advantage — it holds phase state, ticket history, cross-station context, and the loop budget that a single-shot critic lacks — and a decision policy (confidence floor to advance; the round cap; what makes a ticket "86/unrecoverable"). Either extend the critic's output to include a suggested disposition (so the override is a real disagreement) OR rest the proof on "expo refires despite a PASS." Pick one. This converts the pattern from metaphor to spec.

2. Lock ONE canonical exit set

The control-flow exits — the pattern's public API — are enumerated three different ways: advance / refire / 86, then advance / refire / reroute / 86, then a different split. Pick one closed set and use it verbatim everywhere: advance · refire-to-author · reroute-to-spec · 86(kill). Tie each exit to the decision criteria from #1. Present "refire the work vs fix the spec" as the two routing targets, not a separate taxonomy.

Tier 2 — consistency (mechanical, do in the repo reconciliation)

3. Docs-vs-code vocabulary split (unacknowledged)

Prose says brigade/pass/expo/station; the repo still says skill-dev-orchestrator/ and workflow/skill-pipeline-variance-analysis.run.js — the doc literally links a file named "pipeline" one sentence after disowning the word. Decide: adopt fully → rename paths atomically; or keep neutral → add one line mapping (orchestrator = pass/expo, pipeline = brigade). Don't ship two vocabularies in one repo.

4. "86" → kill in the API (founder already agreed)

The mermaid itself translates 86→"Kill" in the same diagram — proving the slang is redundant. Use kill/drop as the primitive; keep "86" as one-line prose flavor only.

5. Verb/noun collisions

• fire (forward) vs refire (backward) — the "re-" silently flips direction; and "advance" is also called "fire," so forward has two names. Reserve one word per direction (e.g. forward = advance; backward = send-back/bounce).
• "fires" also means critic-triggering ("fires on the right asks") — same word, two meanings, same doc.
• pass vs expo used interchangeably for "the decider." Pick one noun (expo = the role inside the pass, per the mermaid label — so say that once and be consistent).

6. Show one real formula (the depth proof shows no depth)

The skill-dev doc's whole thesis is "force depth in," but the variance-analysis worked example is reduced to critic scores (0.78–0.95) and line counts — zero actual managerial-accounting formulas on the page. Put one concrete formula in (e.g. materials price = (AP−SP)×AQ purchased; FOH production-volume = budgeted − applied) so the headline claim is checkable.

Tier 3 — the article needs a different fix than the docs

7. The article wins by strawman (its load-bearing weakness)

It defines "swarm" as pure emergence ("no named step and no one accountable"), so every "swarm structurally cannot" is true by definition. But the most prominent thing named "Swarm" in this space — OpenAI's Swarm framework — is named-agent handoffs, the opposite of anonymous emergence. A knowledgeable reader (exactly the audience that would spread it) takes it down in one line. Related: the "no word yet" hook is false — supervisor / orchestrator-workers / agent-pipeline / blackboard already name this; what the article describes is the supervisor/orchestrator-workers pattern.

Fix: (a) acknowledge "swarm" is contested (also = handoff systems); (b) distinguish "brigade" from the existing orchestrator-worker/supervisor vocabulary — earn the coinage instead of claiming a vocabulary vacuum; (c) soften "structurally cannot" → "the pure-emergent swarm gives this up by design."

8. Price the tradeoff the article hides

It praises the swarm for "no single point of authority," then builds an architecture whose entire value is a single authority (the expo) — a serialization bottleneck and single point of failure. Unaddressed. Name the cost.

Tier 4 — polish (worth a pass, not blocking)

• Undefined-once terms: "depth source," "mission envelope," "the bar" (and given the thesis is who decides, an unowned "the bar" is a real gap — locate it).
• "the pass is the reusable brain" oversells — what's invariant is the loop shape; the judgment policy is domain-specific (it lives in the roster). Downgrade to "reusable control loop," which actually strengthens "thin harness, fat stations."
• Metaphor-mixing for the swappable part: "cartridge" (console) + "fill" + "menu" — pick one; "menu" is kitchen-native and already in the doc.
• Even-handedness undercut: "when a swarm is right" sets up fair, then the doc closes on a brigade-wins mic-drop and runs on a false binary that erases the orchestrated-but-not-a-brigade middle (DAG, supervisor-worker, blackboard).
• Naming collision: "Agent Brigade" is both the pattern and the instance suffix; "Agent" is redundant in every instance name. Consider "{Domain} Brigade" for instances, "Agent Brigade" for the pattern.
• Redundancy: the "return address / send it back" idea is restated ~5×; keep the two strongest. Cut the doc's self-narration of its own honesty ("this is the honest seam," etc.).
• Overstated: "wall-clock = slowest single ticket, not the sum" is true only under unbounded parallelism + no contention; qualify as best case.

The one move if you only make one

Tier 1 #1 — make the expo real. Give it an information advantage + a decision policy. Everything else is consistency and polish; that one is the difference between a sharp metaphor and an actual specification.