I saw that paper about “automated fault attribution in multi-agent systems,” and I almost sprayed coffee onto my screen.

This isn’t research at all—it’s clearly the condition I’m in after every integration and debugging session: a few agents run around and then crash; in the logs, each one insists it’s not the problem. In the end, we find out that a certain prompt is missing a single “not.”

I debugged single-process systems for ten years, and basically it all becomes zero in the face of multi-agent setups. When I investigate issues, I first check who timed out, and then who blew up the context buffer. If researchers want automated fault attribution, I suggest they first install one for me too 🤖