One AI builds. A second AI checks its work. Night one, both found something.
AI agents move fast, and fast work needs checking. So we gave our builder agent an independent checker, whose only job is to test every "done" against what actually happened. On the first night, each found unfinished work in the other's output. It's still running, catching at least one a fortnight before it ships.
Your AI says it's done. How would you know if it wasn't?
What we did, step by step.
Gave the checker a different job
A separate agent, not built to agree with the first one. Its only job is to check. It never builds anything.
Set one rule
Nothing is done until the checker has tested it against the live system. Not against the builder's write-up of what it did.
Ran them against each other on real work
On the first night, the checker caught a fix the builder said was finished that had never actually worked when run. And the builder caught the checker's own health check reporting wrongly.
Proved neither is safe alone
Both gaps were caught that same night, before either reached a person as "finished".
Kept it running
The first few weeks were slow, because tasks were waiting on people. Once the team stayed on it, the checker settled into a rhythm of at least one catch a fortnight.
You wouldn't let a new starter sign off their own work.
The same goes for AI. A builder and a checker want different things, and that's the point.
Wants to finish
It's built to get the job done, so it's inclined to report that the job is done.
Wants to find problems
It's built to test the claim against reality. It gets nothing for agreeing.
From "The AI Says It's Done" to "It's Been Checked"
Taking the AI's word for it
One agent declares a job done. Nobody has a reliable way to check. Gaps surface later, when something breaks.
Every "done" gets tested
A builder and an independent checker. Nothing is marked done until it's tested against the live system. At least one unfinished job caught a fortnight, before it ships.
Check your AI's "done"
Give your AI tool a real task
Something with a result you can check: an email sent, a record updated, a file created.
Check the result, not the report
Go and look at the actual email, record or file, rather than reading what the AI says it did.
Count how often "done" isn't done
Even once is enough to know you need a second check.
Start to finish
"Don't trust the AI that did the work. Trust the one that audits it."
Operations IntelligenceWho's checking your AI's work?
We'll look at how you're using AI today and show you where a second check would catch what slips through.