Stockertim Dictionary · Artificial intelligence
Agent specification
Also called agent spec, agent design document, agent brief
In short
An agent specification is a short written design for an AI agent setting its goal, inputs, allowed tools, step and spend caps, stop conditions, approvals, escalation and what it logs.
- Intermediate
- Document
- Checked
In detail
An agent decides its own steps, so its boundaries have to be decided before it runs. The specification is where that happens. It is usually a page or two, written in plain language, that a builder, a manager and a reviewer can all read and sign off.
A typical specification covers the goal, stated so that success can be checked; the inputs the agent receives; the tools it may use and those it must not; caps on steps, time and spend per run and per day; the conditions on which it stops and hands over; which actions need a person's approval and whose; how it escalates and to whom; and what each log line records. It also names an owner who answers for the agent and a deputy.
The document does real work. Testing is designed against it, because each stated limit can be checked. Reviewers use it to compare the agent's actual configuration with what was approved. When something goes wrong, it shows whether the agent broke its rules or the rules were inadequate.
A specification only matters if the system enforces it. Caps and tool lists belong in code and settings outside the model, not only in the prompt, and any change to the agent's behaviour updates the specification with a version number and date.
Worked example
Rosewood Veterinary Group writes a specification for an agent that books follow-up appointments.
- Goal: for each discharge note, book one follow-up in the right clinic or pass the note to reception with a reason.
- Tools allowed: read discharge notes, search the diary, create a provisional booking. Not allowed: cancel or move bookings, message clients.
- Caps: 8 steps per note and 200 notes a day. Stops on missing pet details or a request outside its listed clinics.
- Approval: reception confirms each provisional booking before the client is told. Owner: head of reception.
Everyone can see what the agent may do, testing has a checklist, and the limits are set in its configuration as well as on paper.
In practice
Teams write an agent specification before building, have it approved by the business owner and a security reviewer, test every limit it states, and keep it under version control alongside the agent's prompts and settings.
Watch out
- Limits written only into the prompt are requests to the model, not controls.
- A specification that is not updated when the agent changes becomes misleading.
Learn it properly
Official sources
- AI Risk Management Framework · National Institute of Standards and Technology
- Guidelines for secure AI system development · National Cyber Security Centre
- AI Playbook for the UK Government · UK Government
Facts last checked on 9 November 2026.
Education, not advice: the Stockertim Dictionary explains what terms mean. It never recommends buying or selling anything.