Cursor 2.4: Subagents, Skills, Image Gen

Subagents run specialized tasks in parallel, Skills enable dynamic workflows, and image generation is now built into agents. Plus Cursor Blame for Enterprise teams.

Cursor 2.4: Subagents, Skills, Image Gen

TL;DR

  • Subagents run specialized tasks in parallel with isolated context — faster, more focused agent work
  • Skills let agents discover and apply domain-specific workflows dynamically
  • Image generation built into the agent — text-to-image with reference uploads
  • Cursor Blame (Enterprise) shows AI vs. human code attribution with conversation links

New

  • Subagents — Independent agents handle discrete subtasks in parallel with custom prompts, tool access, and models. Reduces context bloat and speeds up complex multi-part work.
  • Skills — Define domain-specific workflows in SKILL.md files. Agents discover and apply them dynamically, replacing always-on rules for procedural "how-to" instructions.
  • Image generation — Generate images directly from agent conversations. Text descriptions or reference uploads guide the model; results preview inline and save to assets/ folder.
  • Cursor Blame (Enterprise) — AI attribution in git blame. See which lines came from Tab completions, agent runs, or human edits, with links to the conversation that produced each change.
  • Agent clarification questions — Agents can ask you questions mid-task while continuing to read files, edit, and run commands. Answer arrives and gets incorporated automatically.
  • MCP server optimization — MCP definitions now live as JSON in .cursor. Agents load only when needed, cutting token usage.
  • Browser improvements — 10× faster navigation, drag-and-drop support, improved text input, and browser locking to prevent accidental interference.
  • CLI enhancements — New -continue flag to resume last session, /max-mode toggle, improved vim support, faster warm starts, and unified permissions with editor.
  • Hook improvements — New stop hook, beforeSubmitPrompt, PreToolUse, PostToolUse. Hook commands now 40× faster.
  • PDF support — Agents can now read and attach PDFs as context.
  • Sandbox improvements — Editing skills/rules no longer requires approval, git writes allowed, file edit approvals persist per session, out-of-workspace edits allowed for session.

Fixed

  • MCP server connection performance issue resolved
  • Windows UI text truncation in chat tabs, code blocks, and @-mentions
  • Browser panel rendering on top of other components
  • Browser local network permissions for identity provider auth flows (Okta, etc.)
  • Browser tab splitting and joining
  • Chat tab persistence after reload/restart
  • Syntax highlighting in git worktrees
  • Stale diff views persisting after app restart
  • Tabs disappearing with hidden titlebar
  • Diff UI not appearing for dotfile edits
  • Input lag with rapid text changes in CLI
  • Text wrapping in queued follow-ups in CLI
  • Edit tool usage during plan mode in CLI

Update via: brew upgrade cursor or download from cursor.com/changelog/2-4

Source: Cursor