Cursor Cloud Agents Now Watch Your Slack Threads

Cursor cloud agents can now subscribe to a Slack thread, a pull request or a timed schedule, then wake themselves up when something on that source moves. The change shipped on August 19 in a changelog entry titled "Cloud Agents and Cursor Harness Improvements," and it quietly shifts the product from a tool you steer to a service that keeps going after you shut the laptop.

The centrepiece is subscriptions. An agent binds itself to an event source and stays dormant until that source changes. Cursor says agents now sign themselves up for the pull requests they open and keep pushing until the work is finished — chasing red builds without a human asking twice. For now this only applies to agents running in Cursor's cloud, not the ones running locally on your machine.

Two other additions landed alongside it. A new /goal command hands an agent a single objective it keeps working toward across separate sessions, and subagents can now each be given their own virtual machine. Between them, they answer the oldest complaint about AI coding tools: that the work evaporates the second the chat window closes.

How Cursor Cloud Agents Wake Themselves Up

The idea is closer to a monitoring alert than a chatbot. You point an agent at something that generates events — a PR, a Slack conversation, a recurring schedule — and it sits idle, costing nothing in attention, until that thing changes. A failing check, a reviewer's comment or a Monday morning cron tick becomes the trigger. The most consequential version is automatic: an agent that opens a pull request stays attached to it.

Cloud agents automatically subscribe to PRs they create and drive them to completion.

The Three Features Doing the Real Work

Subscriptions get the attention, but the update only holds together because of what shipped beside them. Persistence, isolation and interruption handling are the three things that usually break when an AI agent is left alone for an hour. Cursor addressed all three in the same release.

/goal, an Objective That Outlives the Chat

Typing /goal gives the agent a long-lived target rather than a task. Cursor's own example is blunt: fix all the flaky tests and get CI green. The agent holds that goal across sessions until it's genuinely done, and it can be combined with a custom mode or with /loop for repeated check-ins. That's a different mental model from prompting — you're setting a condition, not issuing an instruction.

Subagents That Each Get Their Own Machine

Previously, parallel subagents shared a workspace and stepped on each other's edits. Now each one can run in its own cloud environment with a clean copy of the project and a fresh context window. The practical upshot: one subagent can test the parent agent's changes in an untouched environment, or several can swarm unrelated bugs at once without colliding. Isolation is what makes parallelism safe.

Messages That Queue Instead of Interrupting

Cursor also fixed a small but constant irritation. Follow-up messages sent mid-run no longer cut the agent off mid-action; they queue and land at the next tool call. Press Send, or hit Enter twice. Anyone who has derailed a working agent by typing one clarifying sentence will recognise why this matters more than it sounds.

Why Canadian Teams Should Read the Fine Print

"Cloud" is the operative word. These agents don't run on your laptop — they run on Cursor's infrastructure, which means your repository is copied into someone else's virtual machine and left running unattended. For a Toronto fintech, a Montreal health-tech firm or any shop bound by Quebec's Law 25, that's a procurement question before it's a productivity one.

Cost is worth checking too. Cursor's tiers run from roughly US$20 a month for Pro to US$200 for Ultra, with team seats between US$40 and US$120 per user. Billed in American dollars, Pro lands near $28 CAD and Ultra near $280 CAD once conversion and card FX fees are counted. Before enabling always-on agents, three things are worth confirming:

• Where the cloud environments physically run, and whether that satisfies your data-residency commitments
• Which repositories agents may subscribe to, and who can approve a subscription
• How unattended runs are metered, since an agent looping toward a goal burns tokens while you sleep

How This Stacks Up Against Claude Code and Copilot

The autonomy race is crowded. Claude Code runs long multi-file jobs from the terminal with per-subagent model control, GitHub Copilot has moved to issue-to-PR workflows, and OpenAI's Codex is built around extended terminal runs. What Cursor cloud agents add is the waiting part — an agent that does nothing for six hours, then reacts. Adoption suggests the audience is there: JetBrains' January 2026 AI Pulse survey of more than 10,000 professional developers put Cursor and Claude Code as co-leaders around 18 percent work adoption, rising to about 24 percent across the US and Canada.

What to Watch Over the Next Few Weeks

Two signals will tell you whether this is real. First, whether subscriptions come to local agents — Cursor has flagged the cloud-only limit as temporary, and that boundary decides how many teams can use it at all. Second, whether unattended goal-chasing produces merge-ready work or a queue of near-misses someone still has to review. If you're trialling it, start with one low-stakes repo, set a single /goal, and read every PR it opens before you widen the blast radius.