Operating agents in production
The second year is the one that decides whether the first was worth it.
AI agent monitoring is where most of these conversations start and it is never where they end. An agent that has been live six months has a different set of problems from one that has been live six days, and almost none of them are model problems.
What actually breaks
- The upstream changes shape. A model is deprecated, an API adds a required field, a rate limit moves. None of it is your fault and all of it is your outage.
- Credentials expire, usually the ones nobody wrote down.
- The knowledge drifts. The agent is still answering confidently out of a page you deleted in March.
- Cost creeps. Nobody notices a rise in the per-call price until the invoice arrives.
- Permissions granted for a pilot are rarely revisited, and every new tool the agent can call is one more thing it can get wrong at scale.
What AI agent monitoring has to watch
Behaviour, not uptime. An agent returning confident nonsense is up, and every dashboard in the building will agree that it is up.
So the work is: upgrade paths tested before they are applied, credentials rotated on a schedule, a cost line you can read without being an accountant, and a written answer to what the agent can reach and who signed that off.
And a person to call, which is most of what you are actually buying.
Where the knowledge lives
The question that decides whether a deployment survives contact with procurement is where its knowledge lives and whether your own team can read it. If the answer is "inside a vendor's index", you have bought something you cannot audit or leave.
We answer that one first, in writing, before anything gets built.
How it works
Monthly, scoped to a named set of systems, with a response time we agree in advance. It is not a support contract with a ticket queue in front of it. The people who answer are the people who built the thing.