Does an AI coding agent fix its own mistakes?

Rarely on its own. A study of 20,574 real coding-agent sessions found the agent almost never caught a mistake unprompted: a developer had to flag the problem 91.49% of the time before the agent went and fixed it, against just 2.99% where the agent self-corrected with no pushback at all, and 5.52% where the developer ended up fixing it directly instead. The misalignment took seven recurring forms: misreading the project, misreading intent, breaking a stated rule, overstepping its bounds, a bad implementation, a malformed command or tool call, and misreporting its own progress. Most of these episodes, 90.50%, cost extra effort and eroded trust rather than causing damage that couldn’t be undone.

That’s the rate for mistakes a person already caught. What happens to the ones nobody notices is a separate, harder question this study doesn’t answer, since it only counted episodes visible enough to draw a developer’s pushback in the first place.