Used to be, when you asked Cursor to tackle a large feature, you had time to grab two cups of coffee while it worked through everything. That might change now.
Cursor dropped two features today, and they're more interesting together than apart.
Parallel Sub-agents: "Build in Parallel"
Click "Build in Parallel" and Cursor analyzes your task, identifies which sub-tasks are independent, then runs them all at once. Not queuing. Actual concurrency.
Picture this: you ask it to add three new API endpoints, write the corresponding unit tests, and update the frontend calls. Before, it did them one at a time. Now, three sub-agents work simultaneously.
This works especially well for refactoring. When module A changes don't affect module B, parallel execution saves time linearly.
Automatic PR Splitting: From Chaos to Clean Git History
Parallel work is done, now how do you merge? Cursor can now identify logically independent change blocks, propose a split plan for your approval, then auto-create multiple small PRs.
This is way more civilized than the classic "modified 47 files, commit message says update" PR. Each PR does one thing. Reviewers lose fewer brain cells.
Another Signal: /orchestrate
Same day, Cursor shipped the /orchestrate skill — it can recursively spawn sub-agents to handle more complex tasks. They used it internally to research their own skills, cutting token usage by 20% while improving eval scores. Internal backend cold start times dropped 80%.
Recursive agent spawning isn't new, but Cursor packaging it as an out-of-the-box skill lowers the barrier.
My Take
Together, these features push Cursor from "a coding assistant" toward "an engineering team you can delegate to."
Someone summarized it well: Claude Code is a short-lived coding agent with a single async loop. Cursor's path is building a multi-task parallel engineering environment.
What matters in practice:
- Context isolation between sub-agents — if they step on each other, serial might be better
- PR splitting logic — too fine creates PR tsunami, too coarse defeats the purpose
- Token costs — multiple sub-agents running simultaneously, does usage explode?
My gut: small projects won't see much benefit, but for mid-size and larger refactoring, this feature could save real time.
Worth trying "Build in Parallel" before your next big change.
Main sources: