Our view
Agents fail in ways ordinary software does not: they complete the wrong task confidently, loop, or follow an instruction hidden in a document. Those cases have to be hunted, not waited for.
A failure catalogue is also documentation. It tells the team that inherits the system what has already been tried and what it should never be asked to do.
Before you adopt it
- Build a set of known bad inputs and run it on every release.
- Include prompt injection through documents and web pages.
- Log failures with the exact input, model and prompt version.






