Question 1

Does it do real work, on live inputs, for people who did not build it?

Accepted Answer

Yes: Yes. Live inputs, real users, every working day. Partly: Partly. Live inputs, but only the team that built it uses it, or only some days. No: No. It runs on a demo set, a curated sample, or in a notebook.

Question 2

When it fails, is the damage bounded by something other than the prompt?

Accepted Answer

Yes: Yes. A written list of what it must never do, enforced in code, with limits on spend and reach. Partly: Partly. The limits exist but live in the prompt, or cover some actions and not others. No: No. If the model decides to do something, it can.

Question 3

Can someone who was not there read back any run, start to finish, a month later?

Accepted Answer

Yes: Yes. Input, every tool call, every model call and the output, under one id, kept. Partly: Partly. Logs exist, but they are scattered, sampled, or expire. No: No. If it goes wrong, we would have to guess what happened.

Question 4

Is there one named person with a pager and a kill switch?

Accepted Answer

Yes: Yes. One name, reachable out of hours, who can stop it in under a minute. Partly: Partly. A team owns it, or the one person cannot stop it quickly. No: No. Nobody has their name on it.

Question 5

Does it have a budget per run and per month, and does it stop when they disagree?

Accepted Answer

Yes: Yes. A cost per run and per month, enforced, with an alert and a stop. Partly: Partly. We watch the bill, but nothing stops the agent automatically. No: No. It spends what it spends.

Is your agent in production, or in a demo?

Five checks. The ones that decide Tuesday.

Score one agent. Be honest, the card is.