Choose a task someone already owns.
The strongest first agent has a clear input, a clear output, and a person who knows whether the result is useful. Start with a repetitive workflow that creates friction today. Map its exceptions as carefully as its happy path.
Write the evaluation before the prompt.
Collect representative examples. Agree on what a correct answer or action looks like. Include cases where the agent should stop and ask for help. Those examples become the basis for comparing approaches and catching regressions.
Connect the smallest complete loop.
An impressive answer is not a finished workflow. Connect the agent to a source of truth, give it only the tools it needs, and define an approval boundary. Make failures visible and recoverable.
Expand when the evidence supports it.
Measure useful task completion, escalation rates, latency, and cost. Improve the system against that baseline before giving it more scope. The point is a dependable operational capability.