Cursor adds event subscriptions, /goal, isolated subagent VMs, and non-interrupt steering

Cursor is shipping a product-surface upgrade for always-on cloud agents, not a new model. The August 19, 2026 changelog is about the harness: how agents wake up, what they hold onto, where their workers run, and when a human is allowed to nudge them.

The stated aim is blunt. Cursor says it is improving cloud agents and the Cursor harness so always-on agents can operate as a system — building and shipping software without someone stepping in at every loop. In this release, those agents can pick up work from events, hold a goal until it is met, and stay on course through long-running sessions. The official @cursor_ai account posted the same day, highlighting /goal and linking the notes.

That is the right way to read the drop. Long-running coding agents fail less often on raw generation than on the boring systems problems around them: nothing wakes them, the objective evaporates after a few turns, parallel workers collide in the same tree, and a well-meant follow-up aborts the tool call that was about to land. Cursor is packaging those as first-class primitives.

Subscriptions: wake on PRs, Slack, and schedules

Cursor can now monitor pull requests, watch a Slack thread, or run scheduled tasks. The agent subscribes to an event source — a thread or conversation — and wakes when something happens. Subscriptions are available for cloud agents only, for now.

Cloud agents automatically subscribe to PRs they create and then drive those PRs to completion, including fixing CI and addressing bot comments. That is the always-on loop made concrete: open the change, stay attached to it, keep going until the work is actually done rather than dumping a diff and walking away.

The Slack example in the changelog is the same idea in chat form. You can ask @cursor check back in an hour and keep going until that feedback is in. The agent is not a one-shot completion. It is a subscriber that comes back when the thread moves, or when a schedule says it is time to look again.

`/goal` holds the objective until it is done

/goal is the durable objective. You give the agent a long-lived target and it works toward that target until the work is fully complete. The changelog’s example is the kind of ticket people already try to hand an agent and then babysit: /goal fix all flaky tests and make CI green, issued in a new chat.

You can pair /goal with a custom mode so the agent follows a playbook, or with /loop for recurring check-ins. That pairing is the practitioner detail. A goal without a playbook is a wish. A loop without a goal is just a timer. Together they look like a harness: keep this objective alive, use this skill, come back on this cadence.

Custom modes pin a skill as always-on

Custom modes turn any skill into something that stays pinned in the chat. Cursor describes them as “always on” skills. The point is focus — the agent stays on that skill instead of treating it as a one-off invocation you have to re-assert every few turns.

The activation path is in the command menu. From /, pick a skill and press ⌥⏎ on Mac or Alt+Enter on Windows, or choose Use as Mode. Small UX, but it is how a playbook becomes the default posture of a session rather than a reminder sitting in a doc.

Subagents get their own VMs

Subagents can now run on their own virtual machines. Each one gets an isolated copy of the project, clean context, and its own cloud environment.

That isolation is the collision-avoidance story. The changelog points at two patterns: have subagents test the parent agent’s changes in fresh environments, or swarm independent fixes without stepping on each other. The example prompt is run a swarm of subagents to test my app for bugs, each in its own environment.

Anyone who has watched a multi-agent session thrash a shared working tree already knows why this matters. Isolated project copies on separate VMs mean a test swarm does not inherit the parent’s dirty state, and parallel fixes do not collide in the same checkout. The swarm is no longer a set of personalities talking over one filesystem. It is a set of machines.

Steering waits for the next tool call

The last change is quieter and may matter more in long sessions than any of the new slash commands. You can send a message to steer the agent while it is working, without interrupting it. Follow-ups wait for the next tool call instead of cutting the agent off mid-action.

The interaction is simple: type a follow-up and hit Send, or press Return twice.

That is a different contract than stop-listen-restart. Mid-action interrupts waste tool work and scramble the plan the agent was executing. Waiting for the next tool boundary lets you course-correct without tearing down the current step. Combined with /goal and subscriptions, it is how a long-running session stays on course instead of restarting every time a human has an opinion.

What this is — and what the notes do not say

None of this is a model launch. Cursor is packaging event-driven wakeups, durable objectives, isolated subagent compute, always-on skills, and non-interrupt steering as harness behavior for cloud agents. The company says those agents can pick up work from events, hold a goal until it is met, and stay on course through long sessions.

The limits that are actually stated are worth keeping in view. Subscriptions are cloud-agents-only for now. The published notes do not spell out pricing, VM size, concurrency caps, or event sources beyond PRs, Slack threads, and schedules. Those details are unknown from the changelog.

What is specified is the shape of the system: subscribe, wake, hold the goal, spawn isolated workers, accept steering at the next tool call. If you already run cloud agents in Cursor, that is the surface to try.

If you run cloud agents in Cursor, open a new chat and try /goal against a real objective — the changelog’s flaky-tests example is a clean first test — then pin a skill as a custom mode. If you have a PR in flight, let a cloud agent subscribe and drive it. Read the August 19 changelog for the exact commands and the Slack and swarm prompts.