After the launch of Cursor's GitHub version, another major update arrives!
Now, Cursor has given its Cloud Agents a significant upgrade——
Agents can now monitor PRs, Slack threads, and CI pipelines on their own, and can even take on a long-term goal, not stopping until it's done.
Most importantly, /goal is finally live on Cursor.
No Need to Wake It, the Agent Starts Working Actively
The new "Subscriptions" feature allows Cursor Agents to hook onto an event source:
a PR, a Slack thread, a schedule. As soon as there's movement there, it wakes up.
The most powerful part is that cloud Agents will automatically subscribe to the PRs they open themselves and push them all the way to merge.
If the CI fails, they go and fix it; if a bot leaves a comment, they go and reply—no need for a human to open it again.
It's the same in Slack, you can directly tell it, "come back in an hour, wait for feedback, then continue."
However, subscriptions are currently only available for cloud Agents. The local one remains the same, requiring you to speak up for it to act.
One /goal Command, Doesn't Stop Until the Goal is Achieved
The most noteworthy new feature in this update is a command as simple as it gets: /goal
Open a new chat, type /goal fix all flaky tests and make CI green, and then you can walk away.
It will keep working until the goal is genuinely completed.
Put plainly, what you assign changes from a "task" to a "KPI". Before, you said "help me fix this bug"; now, you can say "don't let me see a red light this week."
However, Cursor's own CLI changelog has featured it for a while:
With an active/paused status bar, runs uninterrupted across idle and unattended sessions, Ctrl+C to pause. The official text mentions "rolling out and gated," meaning it's still in a phased rollout.
This wave on August 19th marks its graduation from an experimental CLI feature to being fully integrated with PRs and Slack.
One Isn't Enough, Just Dispatch a Swarm of AIs
The third part: sub-agents can now each launch their own virtual machine.
Each sub-agent gets an independent project copy, a clean context, running in its own cloud environment.
The official usage is straightforward: have sub-agents verify changes made by the main agent in a fresh environment—or have a swarm of sub-agents simultaneously tackle their own problems without interfering with each other.
Previously, it was like a room full of people crowding around the same computer editing the same document—whoever saved last lost work; now, it's one person, one machine, one copy, merging after completion.
Two minor tweaks that will make anyone who's used Agents crack a smile:
Custom modes: Any skill can be pinned as a persistent mode, essentially locking the Agent into a specialized expertise.
Steering: Previously, if you interjected while the Agent was working, it would be interrupted on the spot, like stopping a waiter carrying soup.
Now your follow-up messages take effect the next time it calls a tool—the soup is put down first, then it listens to you.
35% of PRs, All Merged by AI
Just looking at the feature list, many might easily dismiss this as another instance of "AI overpromising."
But in the official blog, Cursor reveals a number——
Among internally merged PRs, over 35% were autonomously created by cloud Agents. This figure was 30% in February 2026, and 0% 18 months ago.
Additionally, starting August 24th, Cursor's Auto will shift to model-based billing.
Included usage for all plans increases; Auto shifts to model-based billing, with quotas adjusted upward accordingly; model routing logic remains unchanged;
Ultra includes SuperGrok Heavy. Takes effect for the current billing cycle, no action required from users.
However, extrapolating from the original Auto mechanism, the direction is clear.
Auto was previously "one flat price per million tokens, regardless of which model was actually routed to." After switching to model-based pricing, that flat-rate pocket narrows.
If routed to a more expensive model, it's billed as such.
While this adjustment in the billing model means the cost of letting these cloud Agents "go all out" will become more realistic and granular.
It's undeniable that Cursor is irreversibly reshaping the entire software engineering pipeline.
From "you type code, it autocompletes" to "you set the KPI, it grinds," even opening its own PRs, running its own CI, and spawning virtual machines to troubleshoot.
This wave of updates from Cursor essentially forces programmers from being "assembly line workers" into becoming "project managers."
Before, opening Cursor meant you were about to start coding.
In the future, opening Cursor might just be to check—how much work did those AIs finish for ByteDance last night.
References:
https://x.com/cursor_ai/status/2090136956101414982?s=20
https://x.com/Da7_Tech/status/2090110167371104398?s=20
This article is from the WeChat public account "New Zhiyuan", author: ASI Apocalypse
