Delegates a task to an agent and checks the result

Using agents · competency using-agents/delegates-and-checks

Taught in: the Using agents course

Draws on: Delegating to an agent, Decomposing work, Working with an assistant

Learning objectives

Writes a brief with goal, context, limits and done-criteria (base)

Claim	Why	Example
A brief states the goal, the context the work needs and the limits it must stay within.	An agent fills in whatever is left unsaid with plausible defaults, so what it must not touch matters as much as what it must do.	"Update the pricing table in docs/pricing.md from this spreadsheet; do not change any other file" stays on task, where "update the pricing" may rewrite the page.
Done-criteria are written before the work starts.	Without them the result is judged by feel, and a plausible result is accepted as a correct one.	"Done when every row in the spreadsheet appears once in the table and the totals match" turns review into a check rather than a read.
The brief names the information the agent should use and where it is.	An agent that has to guess sources picks the most available one, which is rarely the right one.	Pointing at the approved spreadsheet, not "the latest numbers", prevents the agent from pulling last quarter's figures from an old email.

Served by: Breaking a task into pieces an agent can finish, Delegating a task to an agent, Drafting a text and asking for critique, Finding information with an assistant, Thinking a problem through with an assistant, Working from your own files and images

Chooses how much the agent may do before checking in (base)

Claim	Why	Example
The learner sets the check-in point from the cost of a wrong step, however much they trust the tool.	Trust built on a few good runs does not lower the cost of the one bad run.	For a summary of meeting notes the agent may finish unattended. For replies to customers it stops before each send, even though it has drafted well before.
A first run of a new kind of task is done with more check-ins than later runs of the same task.	The first run shows where the agent's defaults differ from the learner's, and that is when the brief gets fixed.	The first time the agent files expense reports the learner reviews each one. After the brief has been corrected twice, they review a sample.
The learner can state in one sentence what the agent may do without asking and what it must ask about.	If the boundary cannot be said plainly, it has not been chosen and the agent will choose it.	"Edit anything under drafts/, ask before touching anything else, never send" is a boundary, and "be careful" is not.

Served by: How much the agent may do before it asks, Delegating a task to an agent

Reviews the result against the brief, not against a feeling (base)

Claim	Why	Example
The review walks the done-criteria one by one and records pass or fail for each.	A result that reads well is accepted by feel, and a checklist catches the criterion the agent skipped.	The brief said "every row appears once and totals match", so the learner counts rows and adds the totals instead of skimming the table.
The review checks the limits as well as the goal, looking for what the agent did that it was not asked to do.	An agent that finishes the task and also "improves" three other things has broken the brief even if the task is done.	The learner opens the list of changed files and questions the two files that the brief did not mention.
A result that fails review sends the learner back to the brief before it sends them back to the agent.	Most misses are something the brief left unsaid, and re-running the same brief reproduces the miss.	The agent used last year's template, so the learner adds "use the template in templates/2026/" to the brief and runs again instead of typing "no, the new one".

Served by: Delegating a task to an agent, Delegating in rounds, Drafting a text and asking for critique

Adjusts the brief when the work reveals new information (base)

Claim	Why	Example
When the agent reports something the brief did not anticipate, the learner updates the brief rather than answering the question in the chat.	An answer in the chat is lost to the next run, but a brief that holds the new information keeps the fix.	The agent asks which of two templates to use, so the learner adds "use templates/2026/" to the brief and restarts from it.
The learner stops the agent when the work shows the goal itself was wrong, instead of letting it finish the wrong thing well.	Finished wrong work costs the run and the review, and a stopped run costs only the run.	The agent's first draft shows the data is quarterly instead of monthly, so the learner stops and rewrites the done-criteria.
A change to the brief is stated as a change, so the agent drops the old constraint.	An agent holds the earlier instruction in context and tries to satisfy both.	"Replace the earlier limit of five bullets with: no limit, one bullet per decision" rather than only stating the new rule.

Served by: Changing the brief while the agent works, Delegating in rounds, When to stop delegating and do it yourself

Alignment

Framework	Code	Asks	Objectives here
Ng, AI engineering skills map	Using coding agents	Plan, execute, verify, monitor at calibrated autonomy	writes-a-brief, chooses-autonomy, reviews-against-brief, adjusts-mid-task
AI Fluency 4D (Dakan and Feller)	Discernment	Judge the output, the process and the behavior of the AI critically	reviews-against-brief
AI Fluency 4D (Dakan and Feller)	Description	State the goal, context and wanted output clearly, and refine it	writes-a-brief, adjusts-mid-task
AI Fluency 4D (Dakan and Feller)	Delegation	Decide what to hand to AI, which tool fits, and how much autonomy to give	chooses-autonomy
Brilliant Coding with AI	SPC-5, BLD-1	Manage constraints; direct an agent to a specification	writes-a-brief
Brilliant Coding with AI	BLD-2, BLD-3	Adjust on new information; divide and delegate	adjusts-mid-task