AI agents are known for making mistakes in multiple ways. Sometimes they just make things up, aka hallucinate. Sometimes they speculate. Sometimes they are unaware of major developments that happened after the model was trained. Sometimes they fail to notice that small things changed while they weren’t looking.

And no, it’s not rare: just today I encountered ALL of the above (examples below). But today I also started getting a new category of excuse: “my subagent did it,” specifically in Cursor multitasking mode.

The subagent did the rebase wrong.
The subagent was supposed to monitor the PR, but somehow didn’t.

Don’t get me wrong: multitasking feels great. But the “subagent” excuse feels kind of lame.

Examples from today:

Hallucination of the day: Claude suggested using OIDC federation that we allegedly already had in our existing workflows, but it couldn’t point to any specific workflow and eventually admitted it was mistaken.

Speculation of the day: “You are getting this bug because you didn’t merge PR #1234.” The thing is, I did merge PR #1234. This one also falls into the “didn’t notice things changed while you weren’t looking” category. Generally, the models in Cursor have a very poor grasp of which PRs were merged and which were not, and they often don’t bother checking.

Missing major development of the day: “No, Linear does not support Releases. You are probably thinking of GitHub Releases.”

I understand that it is impossible to re-check the world every time you say something, but I feel the balance is still off. And the new “my subagent ate my homework” category shifts it even further in the wrong direction.

Mr. Cursor should become responsible for what it has tamed.