Each agent owns one surface. The Marketing Brain owns the whole: an always-on orchestrator that reads signals from every domain, thinks once across all of them, and records what it would do. It's the difference between six specialists optimizing their own corners and one operator seeing that the ad budget question and the discount question are the same question.
The four modes
- off — the default. The Brain doesn't run at all. Nothing changes for existing stores until you turn it on.
- dry_run — the Brain runs its full loop and records every proposal, but nothing executes. This is where you read what it would have done.
- approval_only — proposals land as cards waiting on you. You approve, it executes; you decline, the decline is logged and learned from.
- autopilot — levers that have earned autonomy execute on their own, inside the budget envelope; everything else still comes to you. An autonomy tier (conservative / moderate / aggressive) sets how bold the proposals are.
What a tick does
A tick starts with signal collection — cheap, deterministic reads from every domain. If signals warrant a decision, the Brain assembles a full context bundle: performance and calibration, data freshness, the competitor landscape, the budget envelope, each lever's track record, forecast accuracy, profit economics (MER, CLV), disputes, tracking health, funnel, running experiments, and seasonality. Then it makes one Claude decision pass over the whole bundle and records the resulting proposals. One pass, not one per domain — so the reasoning can trade domains off against each other.
Why a quiet store costs nothing
The expensive part of a tick is the decision pass, and it only runs when signals warrant it. A store where nothing changed — no performance shift, no fresh data worth reacting to, no open initiative needing a step — collects its signals and stops there. No model call, no credits. The Brain scales its cost with how much is actually happening, and the exception cuts the other way: urgent signals like a ROAS collapse or a complaint cluster preempt the normal cadence and get a decision within minutes.
The weekly loops
On top of the tick cadence, a weekly Strategist turns goals into durable initiatives with committed detail plans, and a weekly Reflection distills verified outcomes into at most ten durable lessons the Brain uses as priors. Strategy and learning are scheduled work, not vibes — and both are readable on the Brain page.
Start in dry_run, read a week of proposals, then decide what deserves approval_only. Autopilot is not a setting you pick — it's a state a lever earns.