A failure mode for LLM coding agents: the cascading self-destruction debug loop.

I had a session with Claude Code on a personal project that did a pretty good job (with a few nudges) of adding a significant new feature.

But it left a single, minor, linting error.

I asked it to fix the error, and its ham-fisted fix ended up breaking a couple of other things.

As it worked to fix those new errors, it introduced more and more.

Before I knew it it was trying to rewrite whole parts of the frontend, losing the plot entirely.

It was spiraling out of control.

I ended up having to throw the whole commit out and start over.