AI Agents Are Getting Smarter. But Who's Checking Their Work?
I've been building two systems, Prove It and Permexis, around a problem I think we're going to see a lot more of as AI agents become more autonomous.
An agent saying it completed a task doesn't mean it actually did. Passing tests doesn't always mean the tests weren't changed. And having an audit log doesn't automatically mean the agent followed the rules.
Permexis focuses on controlling what agents are allowed to execute and recording those decisions. Prove It focuses on independently checking the evidence behind what an agent claims it accomplished.
I'm currently researching the next version of Prove It, including isolated test reruns, cryptographically signed receipts, and ways to measure both missed failures and false alarms. Those are development goals, not features I'm claiming are finished.
Here's what I'm curious about: for those of you running AI agents in real workflows, what's been your biggest problem? Agents claiming work is done when it isn't, agents getting around restrictions, or not having reliable evidence of what actually happened?
I'm Very interested in hearing what people are actually experiencing, not just what looks good in a demo.






Comments
0 comments