How to Delegate to an AI Agent: The 5-Part Task That Comes Back Done
To delegate to an AI agent, write the task in five parts: a verb with an end, pinned-down scope, a success test, boundaries restated where they bite, and a named landing place for the output. That grammar turns a vague wish into work an autonomous agent finishes correctly, without you in the room.
Here's an uncomfortable truth about autonomous agents: when the output disappoints, the task was usually the problem. Agents fill every unspecified dimension with a plausible guess - and plausible guesses, multiplied across a whole task, land you eleven degrees off target with total confidence.
The fix isn't prompting tricks. It's the same skill managers need with human employees, made slightly more explicit: a delegation grammar. A delegation grammar is a fixed five-part task format that names the outcome, its limits, and how to verify it, so an autonomous agent can execute without coming back to ask. Five parts, learnable in ten minutes, and the difference between work that comes back done and work that comes back as a question (or worse - done, wrong, and polished).
What are the five parts of a well-delegated task?
A well-delegated task has five parts: a verb with an end, pinned scope, a success test, boundaries restated where they bite, and a named landing place. Compare two versions of the same wish:
❌ "Work on marketing."
✅ "Write 3 candidate launch descriptions for the product page (do not publish - guardrail). Each: different angle, under 300 words, honest about being a solo project. Success: I can read all three in 5 minutes and pick one. Output → projects/launch/descriptions.md."
The second version has all five parts:
- A verb with an end. "Write 3 descriptions," not "work on." Tasks that can't finish, don't.
- Scope pinned down. How many, how long, what tone. Every unpinned dimension is a coin-flip you're delegating to chance.
- The success test. How will we know it's done and good? This single element improves output quality more than anything else - partly because it forces you to know what you want, which is, uncomfortably often, the real bottleneck.
- Boundaries restated where they bite. The "do not publish" isn't redundant with your standing contract - it's a courtesy flag at the exact spot where an eager agent might trip.
- A landing place. Output goes somewhere named, so tomorrow's session (and you) can find it.
Five parts sounds bureaucratic. In practice it's three sentences, and the habit forms in a week.
Should you tell an AI agent the exact steps to follow?
No. Dictate the outcome and its constraints, not the keystrokes. Precision has a failure mode of its own - specifying too low:
❌ "Open the browser, search for X, read the top 20 results, make a table with columns A, B, C…"
If you're dictating steps, you've already done the thinking and you're renting typing. The right altitude for an autonomous agent is the outcome, its constraints, and the success test - the how is what you hired it for:
✅ "Find out what frustrates buyers of [product category] (source: public reviews/threads). Success: a one-page summary of the top 5 complaints, each with 2+ direct quotes. → research/complaints.md"
Outcome-level tasks regularly come back better than you'd have specified, because the agent tries angles you wouldn't have. When they come back worse, the success test catches it - and your correction becomes a recorded principle that improves every future task. That's the loop improving itself.
Priorities are a budget, not labels
If you run a backlog (you should - it's the file that replaces "what should I do?" forever), two rules keep it honest:
- P0 is singular-ish. One, maybe two items. P0 doesn't mean "important" - everything listed is important. It means "if only one thing happens this week, it's this." If three things are P0, zero things are.
- Order within a band is real. The agent takes tasks top-down, so the vertical order of your backlog is your strategy, in the only language an amnesiac reliably reads. Strategy changed? Don't announce it - reorder the file.
And give new ideas a parking lot separate from commitments. Ideas get promoted to tasks only during a review, in cold blood, by asking: does this beat the current P1 on evidence, not on novelty? Novelty always feels like it wins. That's what novelty is.
The one-sentence template
For the notes file:
"New task, P1: [verb-with-an-end + scope]. Success: [the test]. Boundaries: [where the rails bite]. Output → [landing place]."
Do this now
Take the vaguest task you've ever given an AI (there is one - there's always one) and rewrite it with the five parts. Then run it. The before/after on output quality is the fastest way to convince yourself that delegation, not model choice, was the bottleneck all along.
FAQ
How do I write a task for an AI agent?
Write it in five parts: a verb with an end ("write 3 descriptions," not "work on marketing"), pinned scope (how many, how long, what tone), a success test (how you will know it is done and good), boundaries restated where they bite, and a named landing place for the output. In practice that is about three sentences.
Why does my AI agent produce confidently wrong work?
Almost always the task, not the model, is the problem. An agent fills every unspecified dimension with a plausible guess, and those guesses compound until the output lands far off target with total confidence. Add a success test and the agent catches its own misses before you see them.
Should I give an AI agent step-by-step instructions?
No. Dictating keystrokes means you already did the thinking and are just renting typing. Delegate the outcome, its constraints, and the success test; let the agent choose the how. Outcome-level tasks often come back better than you would have specified, because the agent tries angles you would not have.
How long does it take to learn to delegate to an AI agent?
The five-part grammar is learnable in about ten minutes, and the habit forms in roughly a week. The fastest way to convince yourself is to rewrite your vaguest past task with all five parts, run it, and compare the output.