Agents need boring infrastructure

Every agent demo looks the same. A prompt goes in, a tool gets called, a tidy answer comes out, and the room nods. Then someone asks what happens when the API times out, and the room goes quiet.

The gap between a good demo and a useful agent is mostly infrastructure. None of it is exciting, and all of it decides whether the thing survives its first month.

What production actually asks for

1. Identity. Whose permissions does the agent act with? “The service account” is an answer, but rarely a good one.
2. Logging you can read. Not just that a tool was called, but why, with what inputs, and what came back.
3. Retries and timeouts. Downstream systems fail. The agent should fail politely too.
4. An owner. A named person who gets the alert and can switch it off.

If nobody can explain what the agent did yesterday, it isn’t ready to do anything tomorrow.

Start smaller than you think

The most reliable agents I’ve seen do one narrow job, with clear inputs and a human checking the output for the first few weeks. Scope can grow once trust exists. It rarely works the other way round.

The boring work isn’t a tax on the interesting work. It’s what lets the interesting work keep running.