← sources

What’s the worst thing your coding agent has actually done?

source: reddit (reddit:AI_Agents)published 2026-09-25 17:28ingested 2026-09-26 02:35

retained text

Curious what kinds of mistakes people have seen in real workflows.

Not just “it wrote bad code ”, but moments where the agent actually did something you didn’t expect and you had to clean it up after. Maybe it touched the wrong files, ran a command you wish it hadn’t, saw something sensitive, changed too much, called some tool/API in a weird way, or generally had more freedom than you realized.

I’m asking because most agent safety discussions still feel pretty abstract. I’m more interested in the boring practical failures where you suddenly think, “ok, I probably need tighter boundaries before using this seriously.”

Have you had anything like that happen?

And what do you think would have prevented it? Better approvals, sandboxing, read-only defaults, scoped access, better logs, or something else?

document fetched 2026-09-26 00:18 · canonical https://reddit.com/r/AI_Agents/comments/1wq2fdr/whats_the_worst_thing_your_coding_agent_has