Delegating to an agent
Using agents · topic using-agents/delegating
Delegating to an agent is like delegating to a capable new colleague who will fill every gap with a guess. This topic covers writing a brief with a goal, limits and done-criteria, supplying the context the task needs, choosing how far the agent may go before checking in, and reviewing the result against the brief.
Concepts
- Task brief
- The short written statement you hand an agent: the goal, the context it needs, the limits it must stay inside and the criteria that make the work done. Whatever a brief leaves out, the agent fills in with a plausible default, so what it must not touch matters as much as what it must do. Writing done-criteria first turns review into a check. glossary
- Giving context
- Supplying the agent with the information and materials the task depends on: the relevant files, the approved source of numbers, the constraint you know and it cannot, the example of what good looks like. An agent that has to guess sources takes the most available one, not the right one. Point at things by name and location rather than by description. glossary
- Choosing a degree of autonomy
- Deciding, per task, how much the agent may do before you look: draft only, act but ask before anything irreversible, or run to completion. The choice follows from how costly a mistake is and how easily it is undone, not from how capable the agent seems. Start with less autonomy on a new kind of task and widen it as the results earn trust. glossary
- Checking results
- Reviewing what the agent delivered against the brief's done-criteria rather than against a general impression. Read the changes, not the agent's summary of them; run what can be run; look for work outside the limits you set. A plausible result is easy to accept and often subtly wrong, so the brief is the yardstick, and any gap goes back into the next brief. glossary
Links
- Builds on: What an agent is, Verifying outputs
- Leads to: Running a coding agent, Choosing models and tools, Decomposing work
- Competencies drawing on it: Chooses the right tool and autonomy level for a job, Delegates a task to an agent and checks the result
Lessons
- Changing the brief while the agent works (tutorial)
- How much the agent may do before it asks (explanation)
- Delegating a task to an agent (tutorial)
Your reference
Each lesson above adds its takeaways and its example here once you finish it. Your reference lists every lesson you have finished.
Changing the brief while the agent works
Unlocks when you finish Changing the brief while the agent works.
Takeaways
- An agent that asks a question and an agent that guesses are showing you the same gap in the brief.
- An answer in the chat fixes this run. An answer in the brief fixes every later run of this brief, because a new session starts from the files and the earlier chat is not among them.
- State a changed limit as a replacement, and name the old limit. An agent holds the earlier instruction and tries to satisfy both.
- When a first result shows the goal itself was wrong, stop the run and rewrite the goal and the done-criteria. A stopped run costs one run, and a finished wrong result costs the run, the review, and the next run.
- A result comes back, the brief changes, and the run goes again. Steer this run in the chat, and keep the steering in the files.
Example
Do the task in brief.md.
Before I write the update: templates/ holds two files, short-update.md (one paragraph, at most 120 words) and full-update.md (four headings with bullets). The brief says "the template" and doesn't say which. Which one should I use?
How much the agent may do before it asks
Unlocks when you finish How much the agent may do before it asks.
Takeaways
- Most delegated tasks fit one of draft only, act and ask before anything that can't be undone, or run to completion.
- The setting follows from how costly a wrong step is and how easily it is undone. Name the blast radius of the worst step to judge the cost.
- Good results say the agent is capable. They don't change what one wrong step costs, so they don't move the setting past what the cost allows.
- On the first run of a new kind of task, add check-ins such as the plan and the first item, and drop them as results pass review.
- Write the setting as one sentence: what the agent may do on its own, and the step where it stops and asks.
Example
This lesson has no runnable example or prompt block.
Delegating a task to an agent
Unlocks when you finish Delegating a task to an agent.
Takeaways
- A brief has four parts: goal, context, limits, done-criteria. What the agent must not do matters as much as what it must do.
- Name the source the agent should use. An agent that guesses picks the most available source, not the right one.
- Write the done-criteria before sending the task, as things you can tick rather than adjectives.
- 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.
- Review against the brief, criterion by criterion. A result that reads well has told you nothing yet.
- 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.
Example
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]
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.
Sources
DLAI-11Agentic AI: M1 workflows and autonomy, M2 reflection, M4 evals and error analysis, M5 autonomous agents, DeepLearning.AI (course)NgAI engineering skills map: build and deploy AI applications, software engineering fundamentals, use coding agents, shape the build, The Batch (newsletter) (reference)Academy ai-fluency-framework-foundationsAI Fluency: Framework and foundations, Claude Academy (course)Academy introduction-to-claude-coworkIntroduction to Claude Cowork, Claude Academy (course)