Cursor Adds Subscriptions So Agents Wake on Events

Cursor published an update to its cloud agents and to the Cursor harness — the layer the agents run on — on August 19, 2026¹. The company frames the work as ongoing, with the aim of letting agents that are always on behave as a system that builds and ships without someone stepping in at every turn¹.

Three capabilities are named for this release: agents picking up work when an event fires, holding onto a goal until it has been met, and staying on track over sessions that run long¹.

Subscriptions: the agent subscribes to an event source

Subscriptions is the centerpiece. Per the changelog, Cursor can watch a set of pull requests, sit on a Slack thread, or run something on a schedule¹. The agent registers interest in an event source — a thread or a conversation — and starts up when that source produces something¹.

The capability applies to cloud agents at this stage, a limit the changelog marks with the words “for now”¹. Agents running locally are not covered.

The worked example concerns pull requests an agent opens itself. Those get subscribed to without being asked, and the agent stays with each one until it is finished, repairing broken CI and answering what the bots leave behind¹. In Slack, the illustration is telling @cursor to look in again an hour later and to keep at it until the awaited feedback lands¹.

Updates around cloud agents have been arriving steadily. When Cursor introduced builds to Cloud Agents in August so sessions start from a prepared environment, that shortened how long a session took to become usable. What this release adds sits earlier still: the agent subscribing to a source and waking on it¹.

/goal, custom modes, and subagents on their own machines

The other three changes lean the same way — keep the agent running longer.

/goal hands the agent an objective that persists until the work behind it is genuinely done¹. Cursor’s illustration is clearing out flaky tests until CI comes back green, and it suggests combining the command with a custom mode when there is a playbook to follow, or with /loop when periodic check-ins are wanted¹.

Custom modes take a skill and keep it fixed in the chat, which Cursor compares to leaving a skill switched on permanently¹. You select the skill from / and confirm with ⌥⏎ on a Mac or Alt+Enter on Windows, or choose “Use as Mode”¹. Treating a skill as a unit also shows up in GitHub making Agent Plugins 1.0, the shared format that bundles skills and MCP servers, generally available across Copilot surfaces.

Subagents can be given a virtual machine of their own. Each one works from a separate copy of the project, starting with no inherited context, inside a cloud environment reserved for it¹. Cursor suggests two uses: letting a subagent check what the parent agent changed somewhere untouched, and letting several fixes proceed at once without tripping over one another¹. It also sketches turning loose a group of subagents to hunt for bugs, each in its own environment¹.

Then there is steering. A message aimed at an agent mid-task no longer breaks what it is doing; the note is held until the agent reaches its next tool call¹. Typing it and pressing Send works, as does hitting ⏎ twice¹.

It is a smaller change, and one with a different character. The other four extend the stretch of time without a person; this one improves how a person interrupts.

What needs rethinking is permissions and review

Cursor’s announcements over the past few days read as one sequence. On August 17 the company shipped Origin, its code hosting, in early beta, which it described as bringing code, PRs, and agents into the same place. This update touches how the agents there wake and how long they keep going.

For a team adopting this, the impact lands before the question of usability. If an agent is allowed to subscribe to PRs and go fix CI, you need to have decided which branches it can push to, which secrets it can reach, and what it can merge without human approval. If your permissions were shaped around people invoking agents one request at a time, that assumption is worth revisiting.

Review is the same. When an agent keeps making fixes in response to bot comments, you need to decide which point in the diff a human actually reviews, or review turns into a chase.

Running subagents in parallel is also a cost question. Since each holds its own cloud environment, a swarm consumes compute in proportion to the number of environments. The changelog states no pricing or plan availability, so the actual cost has to be checked against your own plan.

One note on framing: the state Cursor describes — agents carrying on without intervention at each turn — is the stated design intent, not something the changelog claims to have reached¹. What this release adds is waking on events and holding a goal.