How do you know what your agents actually did?

At Hyperparam we run agents three ways: coding agents (Claude Code and Codex), a software factory that picks up issues and PRs on its own, and agents that run with no one watching. None of them tells you when something quietly goes wrong, because the models are so good at working around issues. In the first half of September those workers tried to run python3 on a container that doesn't have python3 1044 times, and none of us noticed, because the agents found workarounds and the pull requests looked fine in the morning. We only found out because someone asked.

That's why we built HypAware, launching here on Thursday. We built it because wanted better visibility into our own AI logs. It reads every session from every agent, finds what's broken, proposes the fix, and checks on the next pass whether it held.

What kinds of agents are you running, and how do you find out when one of them goes wrong?