Cursor agents now self-drive their PRs and run on a schedule
Cursor's Cloud Agents update also isolates each subagent on its own VM, pins skills as always-on modes, and queues your follow-ups instead of cutting in.
Copy markdown
Agents subscribe to their own PRs
Cloud Agents now auto-subscribe to the pull requests they open and drive them to completion — pushing fixes and answering review comments without a nudge. They can also watch Slack threads and run on a schedule, so a nightly "triage new issues" or cleanup job kicks off on its own.
One /goal, worked until it's done
The new /goal command hands an agent a long-lived objective and lets it grind toward that goal across many steps instead of one-shot prompts. Point it at a multi-file refactor or "get the suite green" and check back later.
Every subagent gets its own VM
Isolated subagents each spin up on a separate VM with a clean copy of your project, so parallel agents can't stomp on each other's working tree. Fan a change out across packages and merge the results.
Pin a skill as an always-on mode
Custom Modes let you pin a skill so it stays active for the whole session — lock an agent to "only write tests" and it won't drift onto something else. Less babysitting on long runs.
Follow-ups queue instead of derailing
Steering got smarter: a message you send mid-run now waits until the agent's next tool call rather than interrupting the current action, so you can redirect without breaking a working agent.
Elsewhere: Claude designed protein binders for 14 of 15 targets
Anthropic published lab results where a Claude agent ran the full protein-design stack and produced working binders for 14 of 15 targets — a marker of how far agentic tool-use has pushed past code. Anthropic says any lab can now point the agent at the whole pipeline.