Cursor gave its cloud agents a new way to start working: instead of waiting for a person to type a prompt, an agent can now hold a subscription to a pull request, a Slack thread, or a recurring schedule, and it wakes the moment that source changes. Cursor announced the update on its changelog page on Aug. 19. The shift matters because it moves the trigger for agent action outside the chat window entirely.
Cloud agents that open a pull request now attach to it automatically and stay on the case until it merges, rerunning failing checks and responding to bot comments as they come in. A developer can also leave standing instructions in a Slack thread telling the agent to keep monitoring until a specific update lands, and the agent will act on that trigger later without a fresh message from anyone. Subscriptions are limited to cloud agents for now.
Subagents have also moved off the parent agent’s compute. Each one now spins up on its own virtual machine, working from a fresh clone of the codebase instead of sharing memory and file state with the agent that spawned it. That separation lets several subagents run at once: testing the parent’s changes in a clean sandbox, or chasing down unrelated bugs in parallel, without one subagent’s edits stepping on another’s.
Cursor also changed how a person interrupts a working agent. A follow-up message no longer halts the agent mid-action. It now queues behind the current tool call and gets delivered once that step finishes, so a developer can redirect an agent’s approach without resetting its progress.
Together, these changes describe the same underlying shift. An agent that responds to a state change in the world, rather than to a typed instruction, behaves more like infrastructure than a chat interface. That is a meaningfully different failure mode. A prompt-driven agent fails visibly, in front of the person who wrote the prompt.
An event-driven agent that holds a goal across a Slack thread or a stalled pull request for hours can fail quietly instead. Cursor’s post does not say what happens when a subscription goes stale, when an event fires twice, or when the agent misreads an ambiguous trigger. It also does not include failure-rate data or usage numbers for the feature, so the claim that an agent can pursue an objective until it is satisfied rests on Cursor’s own description alone.
The company shipped two related changes in the same release. Custom modes let a person pin a chosen skill for an entire session, and a new goal command gives an agent a standing objective instead of a single task. Both extend the same pattern that subscriptions establish: less prompting, more persistent state carried by the agent itself.
Teams running Cursor agents against live pull requests or Slack channels should treat subscriptions as a new surface to monitor, not just a convenience. A misrouted wake event now changes code without a prompt in the loop to catch it first. Anyone piloting autonomous coding agents this quarter should test how the subscription model handles a stale or duplicate event before pointing it at a production repository.
Cursor detailed these cloud agent and harness updates in its own product changelog, published at cursor.com on Aug. 19, 2026.
