Cursor's cloud agents can now wake themselves up — which quietly breaks how you budget for them

What shipped

The release is titled “Cloud Agents and Cursor Harness Improvements,” and it is five items long. Four of them matter.

Subscriptions. A cloud agent can now monitor pull requests, observe Slack threads, and execute scheduled tasks. Cursor’s description is precise about the mechanism: the agent “subscribes to an event source (a thread or conversation) and wakes when something happens.” The flagship use case is agents automatically tracking pull requests they created through to completion — fixing CI failures and responding to bot feedback without being asked again. Subscriptions are currently limited to cloud agents.

/goal. A new command supplying “a long-lived objective to work towards until it’s fully complete.” The example Cursor chose is /goal fix all flaky tests and make CI green — deliberately open-ended, the kind of task with no natural single-session boundary.

Subagent VMs. Subagents now run on isolated virtual machines, each with an independent copy of the project. The stated purpose is parallel testing and bug-hunting without concurrent operations conflicting.

Steering. You can now send a message to redirect an agent mid-run without interrupting it. Follow-ups queue until the next tool call rather than halting the current action.

There is also a smaller addition — custom modes, which pin any skill as a chat mode to keep the agent focused, reached by selecting a skill from / and pressing ⌥⏎.

The line this crosses

It is easy to read this as an incremental release. Cursor has had background and cloud agents for a while; running work off your local machine is not new.

The new thing is who starts the work.

Every prior version of this feature had a human in the initiating position. You described a task, the agent went away and did it, you came back to a result. The compute moved to the cloud but the trigger stayed with a person. Subscriptions move the trigger too. An agent subscribed to a pull request wakes because CI failed, not because you asked. An agent holding a /goal resumes because the objective is not met yet.

The individual pieces snap together in a way the changelog does not spell out. /goal supplies persistent intent. Subscriptions supply an event stream to act on. Subagent VMs supply the isolation needed to run several of these concurrently without them fighting over the same working tree. Non-interrupting steering supplies the correction channel a human needs when the agent is running work they did not start. That is a coherent design for continuous, unattended operation — and each element is individually modest, which is why the shift is easy to miss.

This is the same movement visible across the category. Warp’s Factories, opened to early access the day before, routes a backlog ticket through fleets of agents for triage, spec, implementation and review. The difference is philosophical: Warp is explicitly model-agnostic and harness-agnostic, letting each stage run Warp Agent, Claude Code or Codex. Cursor is integrating vertically — its own editor, its own cloud agents, and as of the day before this release, its own Git forge. Two opposite bets about where the lock-in lands.

The part nobody costed

Cursor bills on a usage-based credit system. Paid plans map a subscription price to a monthly credit pool — roughly $20, $60 and $200 tiers for individuals, $40 per user for Teams. Auto mode, where Cursor selects the model, is included at no extra cost; manually selecting frontier models draws from the pool, and overages bill at API rates without a markup. Multi-step, multi-file agent operations draw credits for each step and each file processed.

Cursor’s own published guidance is that daily agent users spend $60 to $100 per month, and that anyone running multiple agents should expect $200 or more.

Those numbers were derived from a world where a developer starts each run. Under Subscriptions, the relationship inverts. An agent watching a busy repository wakes on activity in that repository — pushes, CI results, bot comments, review threads. Its consumption tracks how active your codebase is, not how many times your team opened Cursor. A /goal like “make CI green” on a genuinely flaky test suite is, by construction, an objective that keeps generating work until someone fixes the underlying flakiness. That is the point of the feature. It is also an open-ended authorisation to spend.

None of this is hidden or deceptive — Cursor publishes its credit mechanics clearly, and the overage rate carries no penalty markup. But no vendor has yet published cost guidance for event-driven agent usage, because until very recently there was none to observe. The honest position for a buyer is that this is unmeasured, and the way you handle an unmeasured cost is to bound it before you scale it, not after.

The wider pattern is worth naming: the unit of purchase in developer tooling is drifting from a seat to a run. GitHub Copilot moved to token billing for premium requests. Warp bills Factories pay-as-you-go with no subscription. Cursor’s credit pool sits between the two models. In all three, the question “what does this cost per developer per month” is becoming harder to answer, and the answer depends increasingly on machine behaviour rather than human headcount.

Why this matters

Your review process now has a contributor that does not sleep. An agent responding to bot feedback and pushing fixes is a participant in your repository. Branch protection, required reviews and CODEOWNERS were designed for human contributors who make a bounded number of mistakes per day. Before enabling Subscriptions broadly, confirm those rules bind agent-authored commits exactly as they bind everyone else’s — the tooling is not the safeguard here, your repository settings are.

Flaky tests become a spend multiplier, not just an annoyance. Cursor picked fix all flaky tests and make CI green as its own illustration, and it is a well-chosen example of a goal that is easy to state and hard to terminate. If the flakiness is environmental rather than fixable in code, a persistent goal can keep working at it. Scope goals with a definition of done that an agent can actually reach.

Isolated subagent VMs are the least glamorous item and possibly the most important. Running several agents against one working tree produces conflicts that are difficult to diagnose and easy to blame on the model. A clean project copy per subagent removes an entire failure class, and it is the precondition for anything genuinely parallel. It also signals where Cursor expects usage to go: not one agent working harder, but many agents working at once.

The evaluation question is shifting from output quality to operational behaviour. Comparisons like Cursor versus Claude Code or GitHub Copilot versus Cursor have largely turned on code quality and editing experience. Once agents run unattended, the differentiators become how well a tool stops, how clearly it reports, and how predictably it bills. Those are harder to benchmark and matter more. It is the same trajectory as Claude Code’s move to a default auto mode — the interesting decisions increasingly happen without anyone watching.

The verdict

This is a well-designed release solving a real problem. The gap between “the agent finished and opened a PR” and “the PR is merged” is filled with exactly the tedious, interrupt-driven work — CI failures, lint bots, review nits — that developers most resent, and an agent that closes that gap earns its keep. The supporting pieces are thoughtful: isolated VMs and queued steering are the marks of a team that has watched these things fail in practice.

The reservation is not about the feature. It is that the commercial model has not caught up with the capability. Usage-based pricing is fair and legible when a human initiates each unit of usage. When the software initiates its own, the same pricing becomes a variable cost driven by repository activity — and the published guidance every team is budgeting against was measured before that was possible.

Recommendation: Turn Subscriptions on for one repository with a spend cap and a two-week observation window, and pick a repository with ordinary activity rather than your busiest one, so the numbers you collect are representative. Track credit consumption against the same period’s baseline before enabling it team-wide. Use /goal for objectives with a checkable definition of done, and be cautious about goals whose completion depends on conditions outside the codebase. Confirm branch protection applies to agent commits before, not after. If you are still choosing a harness rather than tuning one, the best AI coding tools roundup and our guide for developers are the better starting points — and note that on operational maturity this is now a live differentiator between Cursor and its rivals, not a tie-breaker.

Update, 21 August 2026: the contrast sharpened a day later. Salesforce launched Slack Code, a free multi-vendor container for coding agents that went live with Claude, Devin, GitHub Copilot and Vercel — and without Cursor. Cursor’s absence is consistent with the strategy above: when your product is the environment, plugging into someone else’s room means competing on model output alone, inside walls you do not control.

Questions or corrections? Email Pick Right. Want the full list? See all news.