The mental shift: from prompting to operating
Everything people learn about chatting with AI — phrasing, follow-ups, coaxing — assumes you are in the room. With agents you are not. The agent works between your decisions, so the quality of the outcome is set by what you decide before the run and how you review after it. The middle is the agent’s job.
The working loop has four moves: hire, mission, done-state, review. Everything else in this guide is detail on one of the four.
Hire for a job, not for intelligence
The first operator mistake is hiring “an AI” instead of a role. A role is a one-sentence job description plus a short list of named actions — the shape covered in what is an AI agent. When you pick a role, you are choosing what the agent is allowed to touch, which is a more consequential decision than which model backs it.
Two hiring rules that hold everywhere:
- Prefer the narrow role. A Web Researcher that can only search and read will disappoint you less than a do-everything assistant, because you can tell when it has done its job. (Twelve real examples of narrow roles, with their exact action lists, are in AI agent examples.)
- Separate reading from changing. Give look-things-up work and change-things work to different roles, so a research task can never drift into a config edit.
Write the mission, not the steps
A mission is the unit of agent work: one assignment that either reaches done or fails visibly. The craft is in writing the what and leaving the how to the agent — if you find yourself scripting steps, you wanted an automation, not an agent.
A useful mission has three parts:
- The object. What the work is about, concretely. “Our three named competitors,” not “the market.”
- The deliverable. The artifact you expect: a sourced report, a verdict with the field it was based on, a drafted reply.
- The done-state. The condition under which the mission is finished — including the empty case. “A report on what changed since last week, or an explicit statement that nothing did” is complete; “keep me posted” never ends.
The done-state test
Before handing a mission over, ask: could a stranger look at the output and tell whether the job is done? If not, the agent cannot tell either, and it will fill the gap with confidence instead of an answer. Rewriting the done-state is the highest-leverage edit an operator makes.
Climb the autonomy ladder one rung at a time
Autonomy is not a switch, it is a ladder, and good operators climb it per-job, not per-agent:
- Run it while you watch. First missions run with you looking at the output immediately. You are calibrating: does this role, with these actions, produce work you would use?
- Let it draft; you approve. The agent produces, the output waits in review, nothing moves without your sign-off. Most real agent work should live here — it is where drafting, triage, and reporting belong.
- Put it on a schedule. Only after a mission shape has survived review repeatedly does it earn a schedule: the weekly watch, the morning brief, the standing sweep. Scheduled runs produce the same receipts as manual ones — the review step moves, it never disappears.
The ladder also runs downward, and using it that way is a strength, not a retreat. A scheduled mission whose receipts start needing edits gets demoted to draft-and-approve until the mission is rewritten; a draft that keeps missing gets demoted to run-while-you-watch so you can see where it goes wrong. Autonomy is per-mission and revocable — that is what makes granting it safe.
Skipping rungs is how people end up distrusting agents. Autonomous AI agents that finish the job covers the ladder — and why visibility, not confidence, is what makes the top rung usable.
Review the receipt, not the vibes
Every run should leave a record you can read afterward: what was asked, what the agent did, what came back. Review is fast when you review against the done-state — did the deliverable arrive, are the claims sourced, did it say so where it was unsure?
Two things belong in every review habit:
- Spot-check sources. Follow one or two citations all the way. An agent that links its claims is checkable; checking occasionally is what keeps it honest in practice.
- Read the failures. A failed run with a clear receipt — the lookup missed, the site refused, the answer came back empty — is the system working. Failures that are recorded the same way as successes are what let you trust the successes.
When a run comes back wrong
Diagnose in order — each cause has a different fix:
- Wrong done-state? If the output is confidently beside the point, the mission was ambiguous. Fix the mission, not the agent.
- Wrong role? If the agent lacked the access or the skills the job needed, you hired the wrong title. Move the mission, or split it.
- Wrong rung? If the output needed judgment you did not delegate, it belongs a rung lower — draft-and-approve rather than scheduled.
- Genuinely bad run? Models have bad runs. If the receipt shows a reasonable attempt that went nowhere, rerun before redesigning.
Missions agents do badly
Part of operating well is knowing what not to hand over. Three shapes fail predictably:
- Missions that are really relationships. Negotiating with a client, giving a teammate feedback, anything where the medium is trust — the agent can draft for these, but the mission itself is yours.
- Missions with hidden context. If half the answer lives in your head — politics, history, a promise made in a meeting — the agent will produce something plausible and wrong. Either write the context into the mission or keep the job.
- Missions where variance is unacceptable. If the steps must be identical every single time, you want automation, not judgment — the automation comparison covers when a trigger-action tool is the better hire.
A worked first week
A realistic on-ramp, in the draft-and-approve middle of the ladder: pick the weekly competitor watch from the intelligence use cases. Hire a research role; write the mission — the three competitors, a sourced report, the empty case included; run it once while you watch and mark what you would have wanted different; tighten the done-state; run it again and review the receipt. When two consecutive runs would have been useful without edits, give it the Monday-morning schedule and move your attention to the review step.
That is the entire skill. Not prompt wizardry — a hiring decision, a clear mission, a checkable done-state, and a review habit.