We've been stress-testing AI agents at CrashLabs for months, and one thing is pretty obvious ... every agent breaks in unexpected ways once it meets the real world.

That’s why we built Crashed Out, an open-source library of real AI failures to help builders learn from each other and ship more reliable AI agents.

We’d love feedback from this community:

- What other failures should we add?

- How can we make this most useful to builders?