Marketing Brain

Agents own surfaces; the Brain owns the whole. On each tick it collects signals from every domain, assembles one context bundle, makes a single Claude decision pass over it, and records the resulting proposals as plan_actions rows. One pass instead of one per domain, so the reasoning can trade domains off against each other — the budget question and the discount question are usually the same question.

The pipeline

tick  store=0f8e7d6c-…  mode=approval_only  tier=moderate
  signals → collected from all domains; 2 above threshold
  bundle  → performance · calibration · data freshness · competitors ·
            budget envelope · lever record · forecast accuracy ·
            MER/CLV · disputes · tracking health · funnel ·
            experiments · seasonality
  decide  → 1 Claude pass → proposals recorded (plan_actions)
            google_ads: budget +18% (inside ±20% clamp) → needs approval
            promotions: discount window, 3 SKUs → needs approval
  quiet tick: no signal change → no model call → no credit spend

The signal stage is cheap and deterministic; the decision pass only runs when signals warrant it. A quiet store costs nothing. Urgent signals — a ROAS collapse, a complaint cluster, a reputation crisis — preempt the normal cadence via liveness preemption and get a decision within minutes, and a brain-health monitor repairs stalled loops.

Dispatchable vs proposal-only

• Dispatchable today — Google Ads, catalog, pricing, promotions, collections, social. Proposals in these domains can execute (subject to mode, the trust ramp, and the budget envelope).
• Proposal-only — Meta / TikTok / Pinterest / Bing ads, email, inventory, storefront, feed. The Brain reasons about them with the same bundle, but every write in these domains goes to a human.

The weekly loops

Two slower loops sit above the tick. The Strategist runs weekly and turns goals into durable initiatives with committed detail plans — so the Brain executes a strategy rather than reacting tick by tick. Reflection also runs weekly and distills verified outcomes into at most ten durable lessons, used as priors in future bundles. Learning is bounded and legible: you can read the current lessons, and there are never more than ten.

Per-store modes: off (default) / dry_run / approval_only / autopilot, with autonomy tiers conservative / moderate / aggressive. What autopilot is allowed to touch is governed per lever by the trust ramp.