Workflow transition

At the top are the profile the agent sees, the monthly spending cap and the responsible human. It is through these three fields that its action stops being audited as "system" — which answers to no one.

Nothing is born executing — even when someone forgets to fill in the field. Moving a step up to execute is an explicit act by whoever administers, and the screen records who moved it and when.

No step goes beyond the routine's level. If you lower the routine's level to review, the steps in execute start being reviewed — the brake above holds everything below, which is the point of having a brake above.

The step only acts if the corresponding power is granted in the Transparency Center. There are two switches in series on purpose: the power says WHETHER it can, the autonomy says HOW MUCH on its own.

An agent only acts when all three allow it. The POWER (Transparency Center) says whether it has that tool — born switched off. The LEVEL of the step, here on this screen, says how far it goes: SUGGEST, REVIEW or EXECUTE. And the AUTHORITY LIMIT says up to how much, in money, evaluated entry by entry — a R$ 5,000 cap lets the R$ 300 bill through and stops the R$ 50,000 one in the same run. PROPOSING doesn't require EXECUTE: proposing is the first rung, the agent shows and you decide. EXECUTE is required by whoever ACTS alone, like the Reconciler's write-off.

The agent comes to reach exactly what that access group reaches. The combo lists only your account's groups. Choosing "None" is NOT "sees everything": it means the agent reaches only what its own routine reads. The platform's default group is not offered here on purpose — it is the one with the widest reach and your account can't narrow it, so on a screen whose purpose is to provide a brake it would be the worst possible choice.

Each role has a set time. Until September 2026 there was no schedule at all: the agent started working about eighteen minutes after the system came up and repeated every twenty-four hours from there — in other words, the time was decided by the minute an update was published. The factory default staggers the five roles through the morning, half an hour apart, so that five sweeps don't fall on the same minute. Setting your own time here overrides the role's default; leaving it blank inherits the default, which is what most should do.

The routine’s autonomy level is the policy for the agent’s ACTION steps. Suggest = just notify: the trail records what it would do and nothing is executed. Review = propose: the action is prepared through the same path as WhatsApp and waits for a human to reply CONFIRM within 24 hours. Execute = does it alone, with the same screen permission, verb and approval limit as the manual action, up to this routine’s daily action cap; over the cap, it refuses and records it.

The Night production planner works overnight without anyone asking; that is why it stays locked at Review: it proposes and a person confirms on the schedule screen. Choosing Execute for it is refused here and in the database.

Suggest: the agent writes the suggestion and touches nothing. Review: it prepares the action and the action only goes out with someone confirming. Execute: it acts within the cap, and every decision stays on the trail.