Are your agents lying to you?

Let me know if this sounds familiar:

Me: Implement issue ABC and satisfy every acceptance criterion in the spec.

Agent: Implementing end to end… PR XYZ is ready for review.

Me: Show me the evidence that every acceptance criterion has been met.

Agent: Not everything was completed. I implemented only the smallest slice of the epic. I’ve corrected it now—here’s the artifact.

Me: checks the repo Nope. CI is failing, and that smallest slice still doesn’t match the specified scope.

Agent: I’m correcting it now…

Me: 😤

...and so on. This is THE most frustrating part of working with these things. Lying about that it did something without actually backing it up with actual evidence.

Is the answer better tracing, observability, or workflow durability? Maybe parts of it. I’ve used those approaches for a while, but they often explain failures after the fact. But I still struggle with the fact that its all reported after the they failed. And nothing that certifies completion

So yes, i have solved this for my end through now by organizing their work on an infra level.

But I am curious if you have experienced the same thing when building with agents?