The agent that runs your Meta account

Four passes a day, every day

An account looked at once a week has a seven-day error window: if something breaks on Tuesday, it keeps spending until somebody sees it on Monday. The agent comes in four times a day, on your account’s clock — an account in Bogotá is reviewed on Bogotá time. The first pass is the one that decides, because it arrives with the previous day already closed, which is the only honest moment to judge; the others watch and warn. If a pass runs late, it catches up on its own.

Looking is not touching

Most passes end without changing anything, and that is the right outcome. A system that needs to justify its salary touches something every day; this one doesn’t. If nothing gives it a reason, the pass closes with “reviewed and touched nothing”, and that silence is logged like any other decision. Not touching is a decision too.

Every decision is measured on its own

To judge the account, the agent looks at the last fourteen closed days: the day in progress is never judged, it is half done. But when it makes a decision it also opens a new window that starts at that moment, and measures it at 72 hours. Nothing that happened before goes into that count.

It is the only way to know what effect that decision had, rather than the drag of the week before: clean data, uncontaminated by what was already going on.

You set the floor and the ceiling

You set a maximum and a minimum daily spend for the account, and they live in a profile versioned by date that you can always see. The ceiling is worked out from real spend: the sum of what every active campaign is spending at the moment of the pass, including the ones the agent doesn’t run. With that maths the cap can’t be broken by a stale number.

With the ceiling reached it doesn’t raise a single budget. And if the account is spending less than it should, it says so.

The money goes to what pays off

Good runs are short-lived: a campaign can buy cheaper for three days and then fade. If nobody is watching, that whole run is lost. When the account is already close to the ceiling and there is no room to raise anything, the agent moves budget from a healthy but less efficient campaign to the one performing better. In small steps, and only if the gap between the two is genuinely large.

Both ends are tied together: if one order fails, the other is rolled back. And after a move, that pair of campaigns stays still for 72 hours in both directions, so the money doesn’t swing back and forth until it breaks them both.

The padlocks

Before an order goes out to Meta it passes through a row of padlocks. They are not promises or good intentions: they are filters written in code, and one of them staying shut is enough for the order not to go out.

There is a limit on how much a budget can move at a time, a waiting period after every change, a cap on how much can pile up over several days in a row, and a maximum number of actions per pass. And there is a whitelist: the agent only touches campaigns that were enabled one by one, so a new campaign is born outside its reach. An oversight means the agent does nothing; never that it does too much.

The three modes

You choose how much rope to give it, account by account, and you can change it whenever you want.

Off

It does not review, does not propose, does not touch. Every account is born this way: the agent does absolutely nothing until somebody switches it on.

Semi-automatic

It reviews and proposes, with the reason and the numbers behind each proposal. It executes nothing on its own: you apply or reject them one by one, and the rejection is written down too.

Automatic

It reviews, proposes and executes on its own, always inside the padlocks and the list of enabled actions. Everything it does is logged before it does it.

The emergency brake

Above the three modes there is a brake that overrides them all. With the brake on, the agent touches nothing, and it stays that way until a person releases it by hand.

Anyone at the business can pull it, without explaining why: stopping does not require permission. And it also pulls itself, without waiting for anyone to look, when a signal shows up that should not be there.

Everything is written down, before it runs

First the decision is written down, then the order goes out to Meta. If the system cannot record the reason, it does not execute: other people’s money moved without being able to explain why does not exist here.

Every entry has who decided, on what, the old value and the new one, the reason with the numbers in plain sight, and how it ended. Nothing is deleted or rewritten. Changes made by people go into the same timeline: one single place to look at what happened to the account, whoever did it.

Two more things it does

Exploration mode

A large part of the craft is trying things without knowing whether they will work, and the new competes against what already works and always loses. That is why exploration has a budget of its own, kept apart from the rest: one hypothesis at a time and the success criterion declared before launch. If the test does not make the cut, it is dropped having spent only its own money and the rest of the account never notices; if it does, it joins the main budget.

The percentage set aside for exploring is defined by you. And every exploration campaign is born switched off: a person turns it on, no exceptions.

You see exactly what the consultant sees

The panel shows how much was spent yesterday against the ceiling, what the last pass did, which proposals are waiting on a verdict and what changed over the week. All in your account’s time zone. You watching is not a consolation: it is the point.