Someone built a tool that lets you hear an AI agent narrate its experience as it tries to navigate your codebase. The repository is called Endless Toil. The agent speaks its frustrations aloud while it works.

This is not a productivity tool. It will not make you ship faster. But it might do something more useful: force you to sit with the consequences of your own technical debt.

Most code review catches syntax errors and logic bugs. What it misses is the subjective experience of confusion. A function named `process_data` that actually sends emails. A class with seventeen responsibilities. A module that imports half the application for reasons lost to time. These things pass review because they work. They compile. They ship.

But when an agent tries to understand them, it hesitates. It backtracks. It second-guesses. You hear it happening in real time.

The interesting part is not the AI. The interesting part is that most human developers experience the same confusion but have learned not to voice it. We assume the problem is us. We think we are not smart enough to understand the abstraction. The agent has no such inhibition. It just reports what it sees.

I suspect this will be more useful than linting. Linters enforce style rules. This surfaces cognitive load. If the agent keeps getting confused in the same place, that place is probably confusing. Not wrong. Not broken. Just hard to hold in working memory.

The practical question is whether anyone will actually use it. Watching your code get criticized, even by a machine, is uncomfortable. Most teams already ignore compiler warnings they find inconvenient. Adding an AI that complains out loud might just get muted.

But for solo developers willing to sit through the discomfort, it might work. You cannot argue with a confused agent the way you argue with a coworker. It just keeps being confused until you fix the thing that confuses it.