Cursor’s agent used to make you wait. You’d send a correction mid-generation and it would sit in a queue until the model finished its entire response. Version 1.4 changes that — steering messages now run at the next tool-call boundary, not after the full generation completes.
The key changes:
- Mid-task steering — Messages sent while the agent is working now interrupt at the next natural pause (after a tool call).
⌥+Enter(Mac) /Alt+Enter(Windows) still queues messages for after completion.⌘+Enter/Ctrl+Enterinterrupts immediately. - Usage visibility — A new usage section in settings shows how much of your plan you’ve consumed. No more guessing whether you’re near your limit.
- Improved agent tools — Tool calls are more reliable, though the changelog doesn’t specify which tools or what broke before.
The steerability fix matters more than it sounds. Agent-mode workflows fall apart when you spot a wrong direction early but can’t course-correct without killing the run. Cursor’s previous behavior — queue the message, wait for the full generation — meant you’d watch the agent build on a bad assumption for 30+ seconds before your correction landed. Now the feedback loop tightens to the next tool boundary, which in practice means a few seconds.
The two-shortcut split is a smart design choice. Sometimes you want to interrupt (the agent is heading off a cliff); sometimes you want to append context without breaking its flow. Having both as muscle-memory shortcuts keeps the human in the loop without forcing a binary stop/continue decision.
NOTE
The changelog entry is terse — no version-specific model changes, no mention of context-window adjustments, and no detail on what “improved agent tools” means concretely. Expect more specifics in follow-up patches.
Usage tracking in settings is overdue. Cursor’s pricing tiers gate on fast requests and premium model usage, but until now there was no in-app way to see where you stood. Check it under Settings if you’ve been flying blind on quota.