Eight questions that show
who can own agent-written code.
Ask these before you let anyone, hired or contracted, take ownership of what your agents produce. Fifteen minutes. No trick questions. You are listening for specifics.
The questions
Walk me through the last thing an agent wrote for you that you refused to merge. Why?
A good answer names a concrete flaw: a race, a silent catch, a query in a loop. A weak answer is "it was fine, I merged it."
A feature works on the demo and dies with 10,000 rows. Where do you look first?
You want pagination, memory, N+1 queries, indexes. You want it fast and unprompted.
How do you decide what tests to write for code you did not write?
Listen for "I read it until I can explain it, then test the branches I would be scared of." Not "I ask the agent to write tests."
Production is down at 2am and the code was generated last week. What are your first five minutes?
Logs, last deploy, rollback path, then diagnosis. Someone who has been paged before answers in order.
What does an auditor need from you to sign off on this codebase?
Access control, review trail, secrets handling, who can deploy. If SOC 2 gets a blank look, they have not owned an enterprise system.
Show me a time you deleted more code than you added.
Owners refactor. Generators do not. A real story here is worth more than any certification.
What would you tell a founder who wants to ship agent output straight to prod?
You want a clear no, with a short version of the review they would insist on instead.
Which part of our stack would you want to understand fully before touching anything else?
The right answer is whichever piece money or data flows through. The wrong answer is "wherever you need me."
Anyone who answers seven of these with specifics can own your agent code. That is the bar worth holding. Last page: what it looks like to borrow one for two weeks.
One dedicated engineer, two weeks, free.