Sometimes when I ask the Cursor agent to do something, it starts going through a whole chain of steps—which is great—but midway through, it might do something wrong, or I might think of a change I want to make.
Right now, if that happens, I have to stop everything, tweak my original prompt, and re-run it. But that means the agent has to start over from scratch, even though most of what it already did was fine.
It would be super helpful if we could send in a correction or adjustment while the agent is still working. Just something like: “hey actually, don’t rename that file” or “can you also add types here too?” without restarting the whole process.
This kind of thing wouldn’t matter as much in a normal LLM chat, but in Cursor—where the agent takes big, multi-step actions—it’d make a huge difference for flow and productivity.