Delegating to an AI agent: a practical introduction · Lesson 3 of 5

Lesson 03: Check the Result

Lesson objectives:

Verify an agent's deliverable against the acceptance criteria you wrote earlier.

Spot the difference between "looks good" and "passes the checklist."

Write a three-item verification checklist for a new task.

Prerequisites: Lesson 02 | Previous << 02 Write the ask | Next 04 Spot the mistakes >>

The agent said it was done. That does not mean it is done

Agents are confident. They will say "Here is the completed report" even when the report is missing a column, uses a wrong date, or quietly ignored half of your instructions. The most common mistake is not the agent's error. It is your own failure to check¹².

Verification is not the same as reading the output. Reading is passive. Verification is checking each acceptance criterion against evidence.

Explanation

When you verify, you compare the result to the acceptance criteria you wrote before the agent started. This protects you from two traps: the trap of trusting a confident tone, and the trap of moving on because the output "looks about right." Anthropic's guide on building agents recommends designing clear evaluation criteria and using human feedback at checkpoints; the same principle applies when you are the human².

A simple verification table has three columns: criterion, evidence, and pass or fail.

Criterion	Evidence	Pass/Fail
Every action item has an owner	Read each row and check for a name	?
Every action item has a deadline	Read each row and check for a date	?
Each item is one line only	Count the lines	?

You do not need a fancy tool. A note, a spreadsheet, or a sheet of paper works. The point is that you are checking the agent's output against your own standard, not its own claim.

This diagram shows the loop: the result goes to the criteria, not straight to use. If the result fails, you send it back for a specific fix.

Worked example (follow along)

Priya asked her agent to turn a meeting transcript into a list of action items with owners and deadlines. The agent returned this:

Action	Owner	Deadline
Redesign the sign-up page	Design team	Next sprint
Follow up with the vendor	Priya	ASAP
Review the budget	Finance	15th

Priya's acceptance criteria were: every action item has a specific owner (not a team), every deadline is a calendar date, and every item is one line.

She checks each row:

"Redesign the sign-up page" has "Design team," not a specific owner. Fail.
"Follow up with the vendor" has "ASAP," not a date. Fail.
"Review the budget" has no year. Fail.

Because she wrote the criteria first, she can send back a precise correction: "Replace team owners with individual owners, replace 'ASAP' and 'next sprint' with calendar dates, and add the year to each date."

Your turn (faded example)

You asked an agent to summarize a contract in three sentences. Your acceptance criteria were: exactly three sentences, mentions the price, mentions the duration, and mentions the cancellation clause. The agent returned this summary:

"This contract is for software services. It runs for one year and can be cancelled. The price is competitive and listed in the attached document."

Fill in the verification table.

Criterion	Evidence	Pass/Fail
Exactly three sentences	Count the sentences	__________
Mentions the price	__________	__________
Mentions the duration	"runs for one year"	__________
Mentions the cancellation clause	__________	__________

Answer:

Three sentences: Pass.
Price: Fail ("price is competitive" does not state the price).
Duration: Pass.
Cancellation clause: Fail ("can be cancelled" is vague; the clause itself is not mentioned).

Summary + what's next

Verification is a checklist, not a feeling. The acceptance criteria you wrote earlier become the test. If the result fails, you send it back with a specific correction. In the next lesson, you will learn the four ways an agent is most likely to be wrong, so you know what to look for before you verify.

OpenAI Academy: Agents and Workflows — https://academy.openai.com/public/courses/agents-and-workflows-bieml ↩
Anthropic: Building Effective AI Agents — https://www.anthropic.com/engineering/building-effective-agents ↩ ↩²

Exercises

Take one agent result you received recently. Write a three-criterion verification checklist for it. Check each criterion against the result and record pass or fail. If you do not have a recent result, use the worked example above.

Level 1 (warm-up)

Done criteria · checked locally

Each criterion is specific enough to be checked without guessing.Each criterion has a concrete evidence source (e.g., a line in the output, a date in the document).You record at least one pass and one fail, or explain why you cannot decide.

Before you delegate a task next time, write the acceptance criteria and the verification table first. Then send the task to the agent and fill in the verification table after the result arrives. Note any criterion that was harder to verify than you expected.

Level 2 (advanced)

Done criteria · checked locally

The criteria were written before the result arrived.The verification table has at least three rows with evidence.You note one criterion that was hard to verify and why.

Lesson 03: Check the Result

The agent said it was done. That does not mean it is done

Explanation

Worked example (follow along)

Your turn (faded example)

Summary + what's next

Footnotes

Exercises

Take one agent result you received recently. Write a three-criterion verification checklist for it. Check each criterion against the result and record pass or fail. If you do not have a recent result, use the worked example above.

Before you delegate a task next time, write the acceptance criteria and the verification table first. Then send the task to the agent and fill in the verification table after the result arrives. Note any criterion that was harder to verify than you expected.