The Agent Brigade — Receipts Edition (Lenovo Worked Example, 2026-07-02)
What this is: the 9-step Brigade de Cuisine diagram, with each step annotated by the actual timestamped event from the first full live run. This isn't a proposal diagram — every step below happened, on one day, and the work logs that prove it are filed in the cellar. Times ET.
The compounding chain, in one line: scouting (morning) fed the cellar → the assessment (midday) consumed the scouting → collateral (evening) consumed the assessment — three brigades, one store, zero re-research.
The 9 steps, with receipts
| # | Step (diagram) | Receipt (Lenovo, 2026-07-02) |
|---|---|---|
| 1 | Order submitted to Steward | 12:56pm — "Can you ask the steward to run the assessment brigade for that company?" (founder, iMessage). Earlier that morning, 9:16am, the same pattern scouted Lenovo ("stress test the company research agent brigade"). |
| 2 | Steward gathers context from the Cellar | The assessment ticket's 3 eager sources were already in the cellar from the morning's scouting: canonical identity note, 960-role jobs snapshot, GitHub org signal (58 repos incl. an active genAI tooling cluster). Cellar-first, nothing re-fetched. |
| 3 | Steward writes Ticket, hangs it on the Rail | 1:24:23pm — lenovo-assessment-2026-07-02 enqueued; deterministic contract lint (Gate A) 8/8 pass on first attempt. |
| 4 | Expo checks the Rail | 1:24:34pm — pulled with a 60-minute lease by worker pass-runner-lenovo. Eleven seconds on the queue. |
| 5 | Expo ⇄ Stations delegate-review loop | Step-0 dossier 1:27pm · human gate at C1 1:29pm (approved 1:31pm, delegation recorded in the gate log) · discovery C2 1:44pm (9 candidate work items minted) · classification C3 1:53–2:05pm — the loop's showcase: 9 use-case micro-tickets ran their 10-lens batteries in parallel (90 lens-executions in 12m37s; serial equivalent ≈55min) · prioritization C4 2:22–2:52pm, where the governance earned its keep: the expectation-steward VETOED the first catalog build twice (unlabeled estimate figures; truncation stripping uncertainty language) — refire budget consumed exactly, fixes independently re-verified before anything shipped. |
| 6 | Stations store output in the Cellar once approved | 2:52:19pm — the Build Manifest landed at assessments/lenovo/build-manifest/, schema-validated producer-side. (Later that afternoon a units error found by visual review was corrected by landing a superseding manifest — the original untouched, append-only — and a new deterministic check now makes that error class unlandable.) |
| 7 | Expo updates the Ticket, removes it from the Rail, archives it in the Cellar | 2:52pm — terminal ack: the ticket, now a complete decision trace (every gate verdict, both vetoes, the delegated approval, every artifact ref), filed itself off the rail into the cellar beside the artifacts it produced. One atomic action. |
| 8 | Steward monitors Rail (blocked) + Cellar (completed) | True to design; the one step whose automation is still being wired. On 7/02 a human performed the sweep — including surfacing the evening's 3 collateral drafts paused at the human-release tier (exactly the "blocked tickets" channel the diagram shows). |
| 9 | Steward responds to the requester with the results | 2:52–3:00pm — results + timing table + the manifest's catalog delivered to the requester; by evening the same manifest was rendering in the Use Case Marketplace (9 use-case cards) and had fed the collateral brigade (5 sales assets, 2 published, 3 at the release gate by 7:05pm). |
Why the vetoes are the headline, not a footnote
Three separate judgment vetoes fired across the day's runs — catalog overselling (twice) and a boilerplate objection-sheet answer — and in every case the deterministic gates had already passed the content. The judgment layer caught what rules could not, the refire loop fixed it, and an independent re-check verified the fix before anything shipped. A catalog that cannot oversell and collateral that answers the hard objection honestly is the product; the speed is the bonus.
The numbers
- Assessment end-to-end: 1h 27m wall / ~60m machine (vs a ~6-hour serial baseline)
- Classification phase: 9 UCs × 10 lenses in parallel, 12m37s
- Collateral run: 5 assets in 19 minutes, one real veto → fixed → ratified
- Everything traceable: every dollar figure on every artifact resolves to a named source in the Build Manifest — enforced by a deterministic gate, not a policy memo
What's next (the paces ladder)
Proof of life ≠ proof of scale. In proving order: a real engagement side-by-side against the current ~6-hour manual path (the comparison harness already exists) → a multi-company batch through the gated rails → different intake shapes → a red-team ticket with deliberately planted overselling → and the teachability test that matters most: a colleague drives the steward cold, docs only.
Work-log ground truth: cellar/assessments/lenovo/tickets/lenovo-assessment-2026-07-02.ticket.md (+ the scouting and collateral tickets beside it). Diagram: founder's Miro, v3 2026-07-02.