A person signs off, not another AI.

On our own projects an agent can now write a function, and a second agent can test what the first one wrote. It is quick and it is often correct. A person still carries out the acceptance test before anything is released. I hold that as a rule, and I want to explain why.

Two checks from the same kind of mind

When one AI checks another, it looks like a second opinion. It is closer to asking the same person twice. The two systems were built in similar ways and tend to make similar mistakes. If the first misreads the requirement, the second is likely to accept the misreading.

A person checks differently. They know what the business meant, which is often different from what the specification said.

Who answers for it

The larger reason is accountability. In the industries I work in, banking, telecommunications, oil and gas, and aviation, someone has to answer for every change. A regulator or an auditor will ask who approved it, and they will expect a name.

Above any AI checker there is a named person, and nothing goes out until that person has signed.

So the rule on our teams is short. The last check on any result is made by a person in management, whatever was checked by a machine before it.

Where agents fit

None of this is an argument against agents. They fit best where a role is already well defined. We copy the roles a delivery team already has, analyst, architect, developer and tester, and give each agent the same narrow job a person in that role would have. It works in operations, where the job can be written down. I have not seen it work in marketing, where the job is judgement.

What it costs

A human sign-off is slower, and I accept that. The aim of this work is a better decision, and a result that someone has put their name to is worth more than one that arrived a day earlier.

It is the last line of the position I hold on Data + AI. Organise the data, build the context, connect the model and the agents, and then have a person sign off.