Give your agent a job description

Most agent projects begin with a model and a demo. The ones that reach production begin with something duller: a page that says what the agent is for.

Nobody would give a new hire system access, a budget and a customer list on their first morning without telling them what the job is. Agents get exactly that treatment every day.

Six lines that save a project

A useful agent job description fits on one page:

1. Purpose. The one number it exists to move: quotation turnaround, first-time-right maintenance tickets, days payable outstanding.
2. Scope. The cases it handles, and the cases it does not.
3. Decision rights. What it may approve, change or send on its own, with limits (an amount, a customer tier, a plant).
4. Tools. The systems and actions it may use, each with least-privilege access and every call logged.
5. Hand-over. The conditions under which it must stop and pass the case to a named person.
6. Review. Who checks its work, how often, and against what pass mark.

None of this is technical. All of it is decided by the business, which is why it so often is not decided at all.

Probation, then promotion

Treat the first weeks as a probation period. Start with the agent proposing and a person approving. Measure cost per completed task, error rate and how often a person had to step in. Widen the decision rights only when the numbers earn it.

This is also what makes governance manageable. The EU AI Act, ISO/IEC 42001 and China’s rules for generative AI all ask, in different words, for a clear purpose, human oversight and traceability. A job description, a log and a probation record answer most of those questions before anyone asks them.

The test

If you cannot write the job description, you are not ready to deploy the agent. You are ready for a workshop about which process to start with, which is a perfectly good place to be.