Cursor's new coordinator agent is banned from writing code

Cursor put a coordinator agent into its new Projects beta on September 10, and it never writes code. It plans, starts subagents that run in parallel, and brings their finished work back for review. Cursor sells this inside an editor millions of developers already use, replacing tmux and git worktrees people used to coordinate agents by hand.

The coordinator runs on a dedicated cloud machine and keeps files synced between cloud and local agents. It can also subscribe to a Slack channel, a schedule, or every pull request on a repo. When one of those fires, it decides on its own whether to hand out work.

The no-write rule comes from what goes wrong when people run multi-agent setups by hand. In my personal projects with Claude Code, the runs that go sideways start the same way. A planning session spots a small fix and makes it, and its context fills with implementation detail. It loses the plan and starts reviewing its own edits instead of the subagents'. You can copy the fix without buying anything. Give the orchestrating session a tool list with no write access and make it delegate everything. It feels slow for about a day.

The event triggers worry me more than the delegation model impresses me. A coordinator watching every PR and every Slack message spends money on events nobody chose. Cursor's changelog doesn't mention per-trigger budget caps, and I'd want them before pointing this at a busy channel. Triggers also open a door to prompt injection. Anyone posting in that Slack channel could influence what the agents do.

Others are circling the same problem. Kilo Code 7.6.2 added an experimental shared agent board called Kilo Swarm behind a flag on September 10. The release notes don't say how the board coordinates work. qwen-code 0.23.3 can now hand a subagent turn to Claude Code over the Agent Client Protocol, making it a coordinator for a competitor's agent.

I don't know which of these wins. I'd bet the no-write rule outlasts all three products.