Cursor gave cloud agents subscriptions, and the trigger moved out of the editor

Cursor's 2026-08-19 changelog adds Subscriptions to cloud agents, and it changes what an agent is in a way that a normal feature bullet undersells.

An agent can now subscribe to an event stream and act on it without you prompting anything. Agents automatically subscribe to PRs they open and, in Cursor's own words, "drive them to completion, fixing CI and addressing bot comments." They can also watch a Slack thread. They can run on a cron schedule. (Cursor Changelog)

The same release ships /goal, which holds an agent on a durable objective across turns instead of stopping at the end of a response. Cursor's example is /goal fix all flaky tests and make CI green. The agent keeps working until the condition is satisfied.

And subagents now run on separate VMs with independent copies of the project, so concurrent test runs don't collide over one working tree. That's the hosted version of giving every agent its own git worktree, which people have been assembling by hand with tmux scripts for months.

Three things in one changelog, and together they describe a shape: event-driven trigger, durable objective, isolated execution. That's not an editor feature. That's a job scheduler.

I run a scheduled pipeline every morning and I've felt the difference between prompt-driven and event-driven work directly. Prompt-driven means you're the scheduler. You're the thing that notices CI went red and decides to do something. Event-driven means the noticing is automated, and your job moves to defining the conditions and reviewing the results. That's a genuinely better division of labor for a certain class of work, and a genuinely worse one for another class.

Where it's better: mechanical, verifiable, high-frequency work with a clear success condition. Flaky test fixes. CI green. Dependency bumps. Lint drift. Things where "done" is machine-checkable and the blast radius is bounded.

Where I'd be careful: /goal on anything without a hard stop condition is a budget hazard. "Run until the condition is met" and "the condition is subtly unreachable" produce the same behavior for a long time before you find out, and you're paying for every turn. Cursor's own example, make CI green, is well chosen precisely because CI green is binary and observable. /goal improve test coverage is not that, and the agent will happily grind on it.

The Warp announcement the day before points the same direction. Warp Factories lays out agents across triage, spec, implementation, review and verification, model-agnostic across Codex and Claude Code, with Linear, Jira and Slack integrations. CEO Zach Lloyd said Warp automates about 30 to 35% of its own tasks weekly. (TechCrunch) Vercel put its agent in Slack the same week, read-only by default with plan approval required before any change, which is the more conservative version of the same idea. (Vercel)

Three vendors, one week, all moving the agent trigger from "human types a prompt" to "something happened." The interesting design variable now is what stops it.