Most failed agent projects start too wide. “Handle support.” “Manage deployments.” “Research competitors.” These are not jobs. They are departments compressed into a prompt.
The smallest useful agent has one trigger, one output, and one owner. It does not need a personality. It needs a contract.
An agent that turns a failed CI run into a draft fix is useful. An agent that “improves engineering velocity” is a slide. An agent that summarizes customer invoices with links to the underlying rows is useful. An agent that “understands finance” is a future incident.
The sizing test
Ask three questions:
- Can a human describe done in one sentence?
- Can the agent fail without creating a new emergency?
- Can the output be checked faster than doing the task manually?
If any answer is no, the agent is too large or the workflow is not ready.
The first successful agent in a company should feel almost disappointingly narrow. Good. Narrow means debuggable. Narrow means observable. Narrow means someone knows when it is wrong.
Autonomy is not a feature you add. It is a boundary you earn.