
이미지: X — 테스팅카탈로그
Summary
- Cursor unveiled cloud agent and harness improvements in its August 19 changelog
- The new /goal command lets you assign agents long-term objectives, and subagents now run in their own isolated virtual machines
- Cloud agents can subscribe to PRs, Slack threads, and scheduled tasks, letting them respond without human intervention
- 발표일
- 2026년 8월 19일, 커서 체인지로그
- 신규 명령어
- /goal — 에이전트에 장기 목표를 부여, 완료될 때까지 작업 유지
- 구독 기능
- PR·슬랙 스레드·예약 작업을 감지해 자동 대응 (현재 클라우드 에이전트 전용)
- 서브에이전트
- 각자 독립 가상머신에서 프로젝트 사본을 갖고 실행
- 커스텀 모드
- 스킬을 채팅에 고정해 '상시 실행' 형태로 사용
- 조작 중 개입
- 에이전트 작업을 중단시키지 않고 후속 메시지 전달 가능
Write "check back in an hour" in Slack, and Cursor actually does it
If a Cursor user posts @cursor check back in an hour and keep going until that feedback is in in a Slack thread, the Cursor agent will genuinely reopen that conversation an hour later to check whether the feedback has been addressed. This is the core of "Cloud Agents and Cursor Harness Improvements," the update Cursor posted to its changelog on August 19. Rather than waiting for human input at every loop, the agent has been built to react on its own to events like PRs, Slack messages, and scheduled tasks.
Why this update, and why now
Cursor is an AI coding editor that lets users choose between Anthropic's Claude, OpenAI's GPT, and Google's Gemini, while also training its own model, Composer, in-house. In August 2026, the company was acquired by SpaceX and folded into xAI, and a string of cloud-agent-related announcements has followed since. On the 11th, xAI's new agent "Grok Bot" was rolled out to Cursor Ultra and Team Premium subscribers, introducing a setup where the bot runs its own cloud computer around the clock. On the 13th, it was confirmed that a codebase tab and review tab syncing with GitHub repositories were being tested in a closed beta. This latest changelog appears to be an extension of that trajectory — a reworking of the harness itself so that agents keep tracking state and responding even after a single instruction, rather than stopping there.
/goal hands off objectives and holds the agent to them until they're done
The newly added /goal command gives an agent not a single task, but an objective it must sustain until completion. The example Cursor gave was /goal fix all flaky tests and make CI green. Cursor notes it can be paired with custom modes to follow a defined procedure, or combined with /loop for repeated checks.
Subagents now each get their own machine
Subagents, which the main agent delegates subtasks to, now each run in their own isolated virtual machine. Each subagent gets its own cloud environment with a copy of the project and a clean context, allowing changes made by the parent agent to be tested in a fresh environment, or letting multiple subagents fix issues independently without conflicts. The example command Cursor gave was run a swarm of subagents to test my app for bugs, each in its own environment.
Subscriptions and custom modes
"Subscriptions" let a Cursor agent watch an event source — such as a thread or conversation — and wake up when something happens. This is currently available only for cloud agents, which can automatically subscribe to PRs they've created, fix CI, respond to bot comments, and push the work through to completion. Custom modes pin a specific skill into the chat, something Cursor describes as an "always-on skill." To use it, type / to select a skill, then press Option+Enter on Mac or Alt+Enter on Windows, or choose "Use as Mode."
Lastly, the steering feature lets users send a message to redirect an agent while it's mid-task. Follow-up messages don't interrupt the agent instantly — they wait until the next tool call before taking effect.
Feature overview
| Feature | What changed | Scope |
|---|---|---|
| /goal | Assigns a long-term objective the agent maintains until completion | All agents |
| Subscriptions | Automatically reacts after detecting PRs, Slack threads, or scheduled tasks | Cloud agents only |
| Subagent VMs | Each runs in its own isolated VM with a project copy | Subagents |
| Custom modes | Pins a skill into chat for always-on use | All agents |
| Steering | Sends follow-up messages without interrupting the task | All agents |
How to try it
① Where to start — Type / in the Cursor chat window to bring up the list of skills and commands like /goal and /loop. To connect Slack, simply tag @cursor in a Slack thread.
② Step-by-step usage
- Open a new chat and type an objective statement, such as
/goal fix all flaky tests and make CI green. - Add
/loopalongside it if you need repeated checks. - To keep a specific skill always on, select it with
/, then press Option+Enter on Mac or Alt+Enter on Windows, or click "Use as Mode." - To run multiple subagents, give an instruction like
run a swarm of subagents to test my app for bugs, each in its own environment. - To redirect an agent while it's working, type a follow-up message and click Send, or press Enter twice.
③ Who can use it — Cursor states that subscriptions are currently available only for cloud agents. No separate usage restrictions are specified in the changelog for the other features (/goal, custom modes, subagent VMs, steering).
④ What you can try — For instance, in a repository where CI frequently fails, you could hand off the task of fixing all flaky tests via /goal and move on to other work. When multiple bugs need fixing at once, you can run a swarm of subagents each in its own environment to test in parallel without conflicts. Or you could set up a review request in a team Slack channel and have the agent check in on its own at set intervals.
Editor's take
What makes this changelog interesting isn't that it lists five separate features — it's that they all point in one direction. Handing off objectives (/goal), waking up on events (subscriptions), handling multiple tasks at once (subagent VMs), and allowing course correction mid-task (steering) all add up to a design meant to make agents function not as a single prompt-response loop, but as a system that stays running all day. That intent is right there in the first line of Cursor's changelog: "without needing to intervene at every loop."
Older AI coding tools worked on a back-and-forth model — a person enters a prompt and checks the result. Even after the concept of subagents emerged, they generally shared the same working environment as the parent agent. Giving each subagent its own isolated VM and project copy is a change that will be felt significantly in practice. It structurally reduces problems like multiple agents colliding over the same files, or one agent's failure contaminating other tasks.
For development teams, the subscription feature is probably the most worth testing right away. Instead of a person continuously monitoring a PR until CI passes, you can let the agent respond to bot comments and review feedback on its own, keeping progress moving overnight or on weekends. That said, the fact that this feature is limited to cloud agents means it's still out of reach for teams built around local workflows.
In the coming weeks, it seems likely that Cursor's recently unveiled code hosting product, "Origin," will start to intertwine with these cloud agent features. If a repository lives inside Cursor's own ecosystem, it's a natural next step to run PR subscriptions and subagent VMs without needing GitHub at all. How quickly Cursor plays its cloud infrastructure cards following the SpaceX acquisition will be a good gauge of where this trend is headed.




Comments