Skip to content
AI Training

Delegating a task to an agent

In this lesson we delegate one small task to an agent and get a result we can check. We write the brief and decide how much the agent may do before it comes back to us. Then we review what it produced against the brief rather than against a gut feeling. The task is deliberately small so you can run it as often as you like and throw the results away.

You need any chat-capable agent you already have access to. The lesson works with any product.

The task

Section titled “The task”

Here is the text you hand over. It is fictional; read it once now so that you can judge the agent’s work later.

Memo: office move, second floor to fourth floor

From next month the product team moves from the second floor to the fourth. Desks on the fourth floor are being fitted out over the next three weeks; the move itself is planned for the last Friday of the month, with everyone working from home that day. Facilities will label and transport monitors, chairs, and personal crates. Each person gets one crate and packs it by the Thursday evening before the move. Anything left on a desk after that’s recycled. The printer on the second floor stays where it is; the fourth floor gets a new one in the week after the move. Lockers aren’t moving, so clear yours before the end of the month. Questions go to the office manager, who will hold a short walk-through of the new floor next Tuesday at ten. Attendance is optional but recommended for anyone who keeps equipment at their desk.

The result you want is a one-page summary in this fixed format:

  1. A one-line title.
  2. A bullet each for what’s happening, what you must do, and by when.
  3. One closing line naming who to ask.

Before you write anything, predict: if you paste the memo and type “summarize this”, what do you get back? Probably a paragraph. Possibly a decent one. Unlikely to be in this format, and you have no way to say whether it is right except that it reads well. A brief is how you get the format and the check.

Writing the brief

Section titled “Writing the brief”

A brief has four parts: the goal, the context the work needs, the limits it must stay within, and how you know it is done. The agent fills in whatever you leave unsaid with plausible defaults, so what it must not do matters as much as what it must do. “Summarize this” is a request but not a brief.

Start with the goal and the limits together. The goal is the format above. The limits are the things a helpful agent would otherwise do for you. It might rephrase the dates or add advice the memo didn’t give, or drop the printer because it seemed minor. Write the limits as plain prohibitions. A limit also has to mean the same thing to a second reader. “Keep the printer part short” can mean trim the words or drop the printer, and the agent picks one reading without telling you which. Where the two readings lead to different work, write out the one you mean [1].

Now the context. Tell the agent which information to use and where it is. Here that’s easy: the memo, pasted in full, is the only source. State it anyway. An agent that has to guess its sources picks the most available one, not the right one, and “the latest information” is an invitation to invent.

Last, the done-criteria, and write them before you send the task. Without them you judge the result by feel and accept a plausible summary as a correct one. Good done-criteria are things you can tick: every date in the memo appears in the summary; no fact appears that’s not in the memo; the format has exactly the parts listed.

Checkpoint · multi-choice

Which of these can you tick?

Section titled “Which of these can you tick?”

You are writing the done-criteria for the memo summary. Which of these lines are criteria you can tick?

Select exactly 3.

Put together, the brief looks something like this. Send it to your agent now, with the memo pasted underneath.

Prompt (illustrative, not a recorded transcript)

Summarize the memo below into this exact format: a one-line title; three bullets covering what’s happening, what I must do, and by when; one closing line naming who to ask.

Use only the memo as your source. Don’t add advice, don’t reword dates, and don’t leave out any deadline.

Done when: every date or deadline in the memo appears in the summary, no fact appears that’s not in the memo, and the output has exactly the five lines described.

[memo follows]

Response

Product team moves to the fourth floor next month

  • The product team moves from the second to the fourth floor on the last Friday of the month. Everyone works from home that day.
  • Pack your one crate and clear your locker; anything left on a desk is recycled.
  • Crate packed by the Thursday evening before the move; locker cleared by the end of the month.

Questions go to the office manager, who holds a walk-through next Tuesday at ten.

The response above is illustrative, written to show the structure of a good result rather than recorded from one model, and your result differs in wording. That’s fine, because the brief fixed the structure and the facts rather than the wording.

Checkpoint · repair

Fix the brief

Section titled “Fix the brief”

A colleague sends this to their agent and gets a friendly paragraph with a made-up “see you on the new floor!” at the end. Rewrite the brief so it states goal, context, limits and done-criteria.

Choosing autonomy

Section titled “Choosing autonomy”

A brief says what to do. It doesn’t yet say which actions the agent may take before it checks in with you. For a summary that’s barely a question; the agent writes, you read. For anything that touches the world (sending, deleting, paying, publishing) it is the main question.

That’s the agent’s degree of autonomy, and three settings cover most tasks. At the first, the human decides: the agent gathers or drafts, and you take every action yourself. At the second, the agent proposes and the human approves: it prepares the action and waits for a yes before taking it. At the third, the agent acts and the human reviews after: it takes the action and you check the outcome.

The degree to pick depends on two things: how bad a wrong action is, and how easy it is to undo. Drafting the summary is cheap and reversible, so the agent acts and you review after. Emailing that summary to the whole team is cheap to do but awkward to take back, so the agent proposes and you approve. Deciding which desks people get is a judgment about people, so you decide, with the agent at most laying out the options.

Those two questions set the degree you are working toward. For the first runs of a new kind of task, sit one step below it, and move up once a few results have passed review without surprises. A run that goes wrong one step below costs you one review. The same run at the degree you were aiming for costs you the cleanup [1].

Try it: rerun the task, but this time add one line to the brief that sets the degree. For example, “Draft the summary, then stop and show it to me before doing anything else.” Notice that the agent behaves differently even though the task didn’t change.

Checkpoint · sort

Who decides?

Section titled “Who decides?”

Reviewing against the brief

Section titled “Reviewing against the brief”

You now have a summary in front of you. The temptation is to read it, feel that it is fine, and move on. Do not review by feel. Review by taking the done-criteria you wrote and ticking them one by one.

Go back to your agent’s output and check it against the brief right now: Does it have exactly the five parts? Is the Thursday crate deadline there? The end-of-month locker deadline? The move on the last Friday? The Tuesday walk-through? Is there any sentence that is not in the memo? Write down which criteria fail, if any.

Most agents will pass most of these on the first try. The ones that fail are usually the same ones: a deadline folded into a vaguer phrase, or a helpful addition (“bring a box for anything that does not fit”) that the memo never said. Both read as perfectly reasonable, which is exactly why a review by feeling would have missed them.

When a criterion fails, the fix is not to rewrite the summary yourself. Send the failing criterion back: “The locker deadline is missing; add it and change nothing else.” Then check again. Reviewing against the brief gives you a precise thing to say, and the agent a precise thing to do.

A failure also tells you something about the brief. If the same criterion fails on the second try, the brief left a gap. One agent asks about the gap before it starts, and another fills it with a guess, and either way the gap is yours to close [1]. Close it in the brief itself. The next task then starts from a better one. The brief gives the review something to check, and the review shows what the next brief has to say [2].

Checkpoint · choice

Which review is a review?

Section titled “Which review is a review?”

The agent returns a summary of the memo. Which of these is a review against the brief?

Exercise

Pick a short text of your own that you do not mind sharing with an agent: a newsletter, a notice, a set of meeting notes. Write a brief for a one-page summary of it with all four parts (goal, context, limits, done-criteria), add one line that sets the degree of autonomy, and send it. Then review the result by ticking each done-criterion. A good result is one where you can name every criterion that passed and every one that failed without rereading the whole summary, and where any failure turns into a one-sentence follow-up to the agent.

Stretch: Now delegate the same memo with the wrong format on purpose (a paragraph, no bullets) and see whether your done-criteria still catch a missing deadline.

Recap

  1. A brief has four parts: goal, context, limits, done-criteria. What the agent must not do matters as much as what it must do.
  2. Name the source the agent should use. An agent that guesses picks the most available source, not the right one.
  3. Write the done-criteria before sending the task, as things you can tick rather than adjectives.
  4. Choose the degree of autonomy by how bad a wrong action is and how easily it is undone: human decides, agent proposes and human approves, or agent acts and human reviews after. For the first runs of a new kind of task, sit one step below that degree and move up as reviewed results pass.
  5. Review against the brief, criterion by criterion. A result that reads well has told you nothing yet.
  6. When a criterion fails, send that criterion back rather than fixing the output yourself, and close the gap in the brief before the next task.

You can now

  • Writes a brief with goal, context, limits and done-criteria
  • Chooses how much the agent may do before checking in
  • Reviews the result against the brief, not against a feeling

References

Section titled “References”
  1. Anthropic. Introduction to Claude Cowork. Claude Academy. Course. Academy introduction-to-claude-cowork
  2. Anthropic. AI Fluency: Framework and foundations. Claude Academy. Course. Academy ai-fluency-framework-foundations