An Agent Operating Procedure, or AOP, is the operating context that tells an AI employee how to do its job: the process owner's brief on the desired result, source material, policies, tolerances, approval rules, and known exceptions. It's what a model's general reasoning ability is missing on its own, an organization's specific way of doing the work.

An AOP isn't a static script. When a human reviewer corrects an edge case the agent got wrong, that correction gets folded back into the AOP, so the same mistake doesn't recur. Written in plain operating language rather than code, a process owner can teach an AI employee a new exception without opening an engineering ticket.

As AOPs accumulate across roles, they start to function like a Company Brain: a growing, reviewable record of how the business actually operates.

Frequently Asked Questions

Who writes an AOP?

The process owner, the person who already owns the job the AI employee is taking on, not an engineer. It's written the way you'd brief a new hire: here's the result we need, here are the rules, here's what to do when something doesn't fit the pattern.

How is an AOP different from a static SOP or process document?

A standard operating procedure is written once and goes stale. An AOP is a living document: every reviewed correction updates it, so it reflects how the process actually runs today, not how it was documented a year ago.

Do other vendors use this term?

Yes. Decagon uses "Agent Operating Procedures" for the plain-English instruction sets that drive its customer-service agents. The shared vocabulary reflects a shared underlying idea, though Zamp's AOPs are written for back-office and regulated operations roles (finance, banking, pharma) rather than customer-facing concierge work.

What happens when an AOP doesn't cover a situation?

The agent escalates the case to a person with the evidence and open question attached, rather than guessing. The reviewed resolution then becomes part of the AOP, so the same situation is covered next time.