2026-07-22 · Source: 120-day Jira (~260 tickets) + 16-agent fleet registry, read by Claude subagents;
roadmap from Calvin's AMP pre-read (07-20) + pg-agent-gov AMP site. Items marked "judgment" are Nathan/Claude analysis, not verbatim source.
1Marketing team work patterns
In one sentence: nearly all recurring work is "one event → fan out a full set of tickets across people, channels, and languages".
18 patterns converge into four families; the loop only absorbs this recurring execution — strategy, creativity, and integration are explicitly excluded,
which is the organizational upgrade plan itself (Belle → integrator, Chelsea → CRM planner, Jon → high-value content).
Family
Representative patterns
Trigger event
Who does it
Product launch assembly line
Launch across four lanes (production site / paid LP / PDP+FAQ+manual / SEO), copy production chain (key message → PDP master → ad variants → translation), proofreading tickets, review deployment
New product ships
Belle, Jon, Josh, Francisco, Erica
CRM lifecycle highest volume
Post-purchase sequence (manually built day 0→N on each shipment, ~two weeks of work), welcome series, abandon cart, app update notifications
Shipment / app release
Travis sets strategy, Chelsea executes
Promotional cycle
Monthly promo, Summer Sale audience fan-out (owner / non-owner / upgrade / PA — one ticket each × email/SMS/push), seasonal calendar, flash sale, creator drop (Kiki Wong etc.)
Promo date / new video
Travis, Chelsea, Belle
Platform opsparallel track
Localization handoff (always the second serial step), Amazon listing ops, page maintenance, CRM platform migration (Yotpo → in-house)
Calvin's pre-read (07-20, to Jim · Jeff · Laura) names the program and locks the shape:
12 shared services (中台) every pod uses (registry · chassis · signals · eval · trace · secrets · write-guard · watchdog · console · harvest · data),
7 business pods each with a named taste owner (paid media — Jaime · brand — Laura+Calvin · content — Kelly · CRM — Travis ·
discovery — Josh · web — Erica · ops — Belle+Jenny), and closed loops that ride the services across pods.
Build team: Jeff (owner) · Calvin (architect) · Nathan (tech lead) · Erica & team (builders) · Belle (deployment).
Keystone test for every component: if a 2× stronger model ships tomorrow, does this get MORE valuable — or worthless?
P1·pre
Spontaneous agents 9 built, 8 live ✓
Phase 1
Now → Q3 · eval spine on paid-media, taste oracle shared, Loop 1 closes
2027 H1 · data service, first T3 autonomous graduation
Trust ladder: T0 sandbox → T1 shadow → T2 production (eval-gated) → T3 autonomous.
Graduation arithmetic (ported from ERP, already live there): ≥95% match × 2 months × 0 uncorrected interventions × named sign-offs.
Registry today: 16 agents, 8 live in production, zero at T1 or above; per-phase ROI checkpoints owned by Jim (COO).
The keystone gap is unchanged — fleet-wide eval spine is still scaffold-only (no golden sets, no calibrated judges, scores null).
But it is no longer theory that the loop can close here: two loops have now closed —
Reactor visual-codex (curator agreement 30% → 80%) and Calvin's own taste loop (first cycle 07-17, below).
pg-agent-marketing Phase 0 (contracts + judge + calibration + outcome write-back, built 07-16) is the machinery that
generalizes exactly this pattern to every pod — same structural fix as ERP's graduation-evidence mechanism.
2bThe first taste loop has turned — cmo-agent, cycle 1
On 07-17 Calvin ran the first full cycle of the taste loop (CMO-001): real creative verdicts, captured and distilled
into machine-readable doctrine that every future session loads. This is the seed of the shared taste oracle that
Phase 1 graduates to Laura co-ownership — the "definition of good" every pod's output will be judged against.
taste-doctrine v1.1verdicts become standing rules with sources
▼
Every next sessiondoctrine + lessons auto-loaded — the agent judges with Calvin's taste
every verdict sharpens the oracle
What cycle 1 actually distilled (examples from the ledger): the word "AI" is banned in customer-facing copy — brand the tech
as Amp Intelligence™; DJI-level terseness is the standing direction (mechanism → result, one line per section, visual carries emotion);
Spark-adjacent language on REACTOR is a kill signal; layered headlines — chapter headers may carry emotion, section headlines must be functional.
Why this matters (judgment): the loop shape works with n=1 human. The scaling question is no longer "can taste be encoded?" —
it is "who else's verdicts feed the ledger?" That is what each pod owner's golden set (≈30 judged examples) answers.
3Platform architecture — who lives where
Each layer is a place. Read top-down: people touch only the top layer; everything below is machinery.
People — Dashboard & Slack
marketing team · Google sign-in · no GitHub accounts
ApprovalsLabeling & feedbackStatus boardSlack bot
App plane — Supabase
runtime state · role permissions · every click audited
Commands queueApproval queueEvent logCalibrationCampaign status
Honest note: what exists today is the judgment machinery (operator repo, 44 tests) and marketing-pm; the dashboard, Supabase schema, runner, and BI hookup are design, not yet built.
Humans appear in exactly one box. Everything else is machinery — and the return arrow is the whole point:
every measured result makes the next judgment better, so the human box keeps shrinking.