Agents That Check Each Other's Homework

Never let the same student write the exam and mark it.

You'll get 100% every time.

The student and the teacher

In school, you write the exam.

The teacher marks it.

Why?

Because when you check your own work, you see what you meant to write, not what you actually wrote.

AI has the same blind spot.

An agent that wrote the code is often too confident about it.

The builder and the reviewer

So I split the roles.

Builder agent: writes the feature.

Reviewer agent: a separate agent, with fresh context, asked to find problems.

The reviewer hasn't "fallen in love" with the code.

It just reads it.

And it's surprisingly good at spotting what the builder missed.

Briefing the reviewer

Bad:
Is this code okay?
 
Good:
You are reviewing a pull request you did not write.
Look for: bugs, missing edge cases, security issues,
and anything that breaks existing patterns.
List each problem with the file and line.
Do not rewrite the code.

"Is this code okay?" almost always gets "Looks great!"

That's like asking your mother if your drawing is good.

Tests are the external examiner

Agent reviews are helpful.

But tests are the board exam.

They don't have opinions.

They pass or they fail.

So my flow looks like this:

1. Builder writes the code and tests.
2. Tests run.
3. Reviewer agent checks the changes.
4. I review the final result.

Three layers before anything reaches the main branch.

Meet Suman, the QA

Suman is a QA engineer.

She uses a reviewer agent to suggest edge cases she might not have thought of:

"What happens with an empty cart?"

"What if the network drops mid-payment?"

She still decides what to test.

AI just widens her checklist.

Final thoughts

The writer should never be the only marker.

Not for students.

Not for developers.

Not for AI.

Give your agents a teacher.

Give your code real tests.

And save the 100% for work that actually earned it.