Cursor just turned its cloud agents from assistants into employees who never clock out. A new /goal command turns a single prompt into an end-to-end mission - "fix all flaky tests and make CI green" - and the agent keeps working, overnight if needed, until the goal is actually met. Underneath the feature list sits the bigger signal: over 35% of Cursor's internally merged pull requests are now created autonomously by cloud agents, up from 30% in February 2026 and 0% eighteen months ago. The software engineering pipeline is being rebuilt around agents that do not just autocomplete code, but own the entire job.
From "Fix This Bug" to "Make CI Green"
Start a fresh conversation, type /goal fix all flaky tests and make CI green, and walk away. That is the pitch, and it changes what you are actually delegating: not a task, but an outcome. Previously you told an agent "fix this bug." Now you tell it "I do not want to see red lights this week." The command had already existed inside Cursor's CLI - marked "rolling out and gated," running unattended across idle periods without interruption, with an active/paused status bar. On August 19 it graduated from experimental CLI feature to first-class product, wired directly into pull requests and Slack.
Agents That Wake Up on Their Own
New Subscriptions let an agent attach itself to an event source - a pull request, a Slack thread, a scheduled job - and wake up the moment something happens. The aggressive part: a cloud agent automatically subscribes to its own PRs and pushes them all the way to merge. CI breaks, it fixes it. A bot leaves a comment, it replies. In Slack you can tell it "come back in an hour, and continue when the feedback arrives." No human needs to open a single view. The one caveat: subscriptions are cloud-only for now. Your local agent still waits to be asked.
A Swarm of Subagents, Each With Its Own Machine
The third piece is about scale. Subagents now get their own virtual machines - each with an independent copy of the project and a clean context, running in its own cloud environment. The official use case is verification: have a subagent validate in a fresh environment what the main agent changed. But it scales to "a swarm of subagents, each fixing its own problem in isolation, without colliding." The old model was a room of people editing one document on one computer, where whoever saved last won. The new model is one machine per agent, merge when done. Two smaller changes are worth noting: Custom modes let you pin any skill as a persistent mode, effectively locking an agent into a specialty, and Steering means your follow-up messages take effect at the agent's next tool call instead of interrupting it mid-task - like a waiter putting the tray down before taking your order.
35% of Merged PRs Are Already AI-Created
Cursor backed the features with a number: more than 35% of the PRs internally merged at the company are now created autonomously by cloud agents. That was 30% in February 2026 - and 0% eighteen months ago. The billing change is the quieter but sharper signal. Starting August 24, Cursor's Auto tier moves from a flat per-million-token price to per-model pricing, with included usage raised across all plans (Ultra picks up SuperGrok Heavy). Under the old Auto mechanism, the same token price applied no matter which model got routed; per-model pricing closes that flat-rate pocket, so routing to a pricier model now costs more. Making agents grind overnight is about to get a real, itemized price tag.
Programmers Become Project Managers
This is not another round of AI hype. It is a rearrangement of who does what on a software team. Eighteen months ago, zero Cursor PRs were AI-created; now more than a third are, and the trajectory does not bend back. The programmer's role is migrating from assembly-line worker to project manager - you no longer write every line, you set objectives and review what a swarm of agents produced overnight. The shape of the shift is identical to the one that made code hosting go agent-first with Cursor Origin: the IDE stopped being a text editor and became a factory floor.
What Your Team Should Do Next
First, read the 35% figure as the trend line, not the headline - your pipeline will look like this before the year is out, so build review and acceptance gates now. Second, start with bounded goals: give a cloud agent one measurable /goal with a clear definition of done, and inspect the diff before merge. If you want to practice the review workflow cheaply, the free tiers of mainstream AI coding IDEs like Trae are a low-risk place to start. Third, respect the local/cloud split: subscriptions and long-running goals are cloud-only, and the shift to per-model billing makes agent overtime a real line item - so define what "done" means before you let one grind. The team that masters reviewing and accepting agent output, not the team that writes the most prompts, will win this transition.
Frequently Asked Questions
Is /goal available for local (non-cloud) agents? No. Subscriptions and long-running /goal missions are cloud-only for now; your local agent still starts each task when you ask. The /goal command is also rolling out in a gated fashion, so not every account sees it immediately.
How much of Cursor's own work is done by agents? Over 35% of the pull requests internally merged at Cursor are now created autonomously by cloud agents, up from 30% in February 2026 and 0% eighteen months ago.
Will autonomous agents replace programmers? The evidence points to a role change, not a disappearance. Programmers shift from writing every line to setting outcomes, reviewing diffs, and owning acceptance - the same shift /goal brings to Cursor itself.
