The short version
- An SOP assumes a person is reading it and can fill the gaps with judgment. An AI agent cannot.
- An AOP, or Agent Operating Procedure, is a procedure written so software can follow it step by step, with the judgment calls written down.
- The term is early and partly vendor-driven. Decagon pushes it most. Skan AI and AthenaIntel use it too.
- To write one, start from the SOP you have: state the goal, list the decision points, define the actions and tools, set guardrails, then test.
The procedure your business runs on was written for a person
A standard operating procedure assumes a human is reading it. It assumes that person can interpret a vague step, ask a coworker when something is unclear, and use judgment to fill the gaps the document leaves open. That is fine when a person runs the process. It falls apart the moment you hand the same document to an AI agent and expect it to act.
An agent cannot run an SOP the way a person does. It needs the intent stated, the decision points named, and the actions made explicit. The procedure has to be written for the thing that will carry it out. For a growing share of routine work, that thing is now software.
What an AOP is
An AOP, or Agent Operating Procedure, is the agent-era successor to the SOP. Where an SOP tells a human how to run a process, an AOP is a procedure written so software can follow it step by step, with the judgment calls written down. It captures the intent, the decision points and the actions, things like book a job, check a warranty or update a record, rather than only describing them in prose.
It is an emerging term, pushed mainly by Decagon and also used by Skan AI and AthenaIntel. It is not yet a universal standard, but it names something real that is showing up across the AI space: the agent-runnable version of a procedure.
| Compared | SOP | AOP |
|---|---|---|
| Written for | A person | An AI agent |
| Gaps are filled by | Judgment, or asking a coworker | Decision points stated in advance |
| Becomes | A document people read | Steps software can follow |
| Exceptions | Handled as they come up | Handed to a person by design |
An example: booking a service call
Each step is a stated decision or action, rather than a sentence a person has to interpret:
- Confirm the address is inside the service area.
- Check the customer record for open jobs or warranties.
- Offer the next two open slots for that job type.
- Book it and send the confirmation text.
- Log it in the job system.
Anything outside the rules, such as an emergency, a complaint or a price question, goes to a person.
How to create an AOP
You do not start from scratch. You start from the SOP you already have and make it explicit enough for an agent to run. The steps:
- Start from the SOP. The written procedure your team already follows is your source of truth.
- State the goal. Name what a successful run looks like in one sentence, for example "book the customer into the right slot for their job."
- List the decision points. Write out every place a person would normally pause and judge: the checks, the thresholds, the if-this-then-that. An agent cannot guess these the way a person can.
- Define the actions and tools. Spell out the concrete actions and the systems they touch: look up the customer, check the calendar, book the job, send the confirmation.
- Set the guardrails. Decide what the agent must never do on its own and where it has to hand off to a person.
- Test and validate. Run it on real cases, compare against how a person handled them, and tighten the steps until it holds.
In short: an SOP describes the work for a person, and an AOP makes the same work runnable by an agent, with the judgment written down instead of assumed.
Where it stands
AOP is early. The term is partly vendor-driven, and full autonomous agent operation is still a direction rather than a finished product you can buy today. The practical move is to turn the SOPs you already have into agent-runnable procedures, start with the routine work, automate what is safe to automate, and keep a person on the hard cases. The agent earns more of the work as it proves it can run the work.
That is the difference between chasing a trend and building on one. The trend is agents. The thing you build is the procedure they run.
Common questions
What is an AOP?
An AOP, or Agent Operating Procedure, is a procedure written for an AI agent rather than a person. It is the agent-era successor to the SOP: a procedure written so software can follow it step by step, with the judgment calls written down. It is an emerging term, pushed mainly by Decagon and also used by Skan AI and AthenaIntel, and is not yet a universal standard.
Do I need to throw out my SOPs?
No. Your SOPs are the source. An AOP starts from the procedure you already have and makes the goal, the decision points and the actions explicit enough for an agent to run. Your team keeps the SOP. The agent runs the AOP.
Is AOP a real standard yet?
Not a universal one. AOP is an emerging, partly vendor-driven term as the market shifts from workflows to agents. It names something real, the agent-runnable version of a procedure, and full autonomous agent operation is still a direction, not a finished product.
Are AOPs the same as agent prompts or workflows?
Not quite. A prompt is a single instruction and a workflow is a fixed sequence of steps. An AOP sits between them: a written procedure with its decision points spelled out, so an agent can handle the branches a rigid workflow cannot, and stay consistent in a way a one-off prompt cannot.
If you want help
Turning SOPs into procedures an agent can run is part of Build, our custom agents and automations work, scoped after a Head Start or for members. Bring one process your team runs by hand to a free 20-minute call.
Sources
- Decagon: Agent Operating Procedures, from manual SOPs to automated AI logic
- Decagon: Agent Operating Procedures (product page)
- Skan AI: why AI agents need more than standard procedures
- AthenaIntel: Agent Operating Procedures (documentation)
Update log
- Moved to the new Signals layout and edited. Added a worked example, replaced a broken Decagon link, pointed the Skan AI and AthenaIntel sources at the pages that use the term, and cut the questions that repeated the post. The definitions are unchanged.
- First posted.
Current as of June 27, 2026. Signals is general information, not legal, security or financial advice. How we source Signals.
