Developer tools: autonomous coding agents
Cursor wants to run your agents and host the code they write
Cursor shipped Projects and Origin, a code host that can be the source of truth for repositories created inside it. Both arrive in beta, and the gaps in the announcement matter as much as the features.
Emmanuel Fabrice Omgbwa Yasse AI-assisted
2026-09-26 · 4 min read

Cursor's release notes usually read like feature lists. This one reads like a claim on territory. In an announcement published today, the company launched Projects, a layer where a coordinator agent plans work and hands it to subagents, alongside Origin, a code host of its own with repositories, pull requests and two-way GitHub sync. Cursor has been climbing toward owning more of that stack, trading independence for the world's largest GPU fleet on the way. Origin is the piece it did not already control.
Both arrive in beta rather than as finished products, and the questions they leave open matter more than the feature list.
What Cursor actually shipped today
Projects handles the bigger jobs: a feature, a migration, a whole application. Its coordinator agent does not write code itself. It plans the work, delegates it to agents that implement, then returns finished output for review, running as many agents in parallel as the task requires. Cursor has already shipped pieces of that pattern, including parallel chats for side explorations. A project runs on its own machine in the cloud, so closing a laptop does not stop it, and when a test has to run locally, the coordinator spins up a local agent to run it there.
Origin covers hosting, starting with repositories, pull requests, code navigation and GitHub sync, with agent-native features promised later. Cloud agents no longer need a GitHub or third-party SCM connection to start, and self-hosted workers now support computer use on Linux and Mac.
Projects is in beta and rolling out gradually to all users from today. Origin reaches early beta on paid plans today, excluding Enterprise organizations whose admins opt out.
Origin vs. GitHub: the source-of-truth question
Cursor does not pitch Origin as a GitHub replacement, and the mechanics back that up. For repositories hosted on Origin, Origin is the source of truth and pushes never reach GitHub. For repositories imported from GitHub, GitHub remains the source of truth, pushes continue to go there, and Origin mirrors the result. Pull requests on synced repositories move both ways: a comment left in Cursor posts to GitHub, and a reply on GitHub lands in Cursor within seconds.
The weight sits in the other case: a repository created in Origin, worked on by agents, reviewed and merged inside Cursor, never touching GitHub. Repositories carry a codebase name that becomes part of their URL, cursor.com/codebase/acme-corp, and Cursor is wiring in an app ecosystem alongside, with Vercel, Depot and Buildkite integrations available now and Vercel giving every pull request a preview deployment.
The coordinator-and-subagent claim, tested against the wording
The claim to isolate is that the coordinator does not write the code. It plans, delegates and returns work for verification, running as many agents in parallel as the work demands. Cursor says Projects delegates tasks to thousands of subagents. Subagents can now run on their own virtual machines, each with an isolated copy of the project and clean context in its own cloud environment, which removes the risk of parallel patches colliding in the same files. A fresh window per subagent also answers something research on dead agent context keeps pointing at: the tokens an agent spends on material that stopped being useful several steps earlier.
That is a sensible answer to a real problem. What the announcement does not offer is a measurement. No success rate, no ceiling on how many agents one coordinator manages before quality drops, no published example of a project running for the months of context Cursor describes. A Project keeps a shared file set synced across its cloud and local machines, and agents add research, artifacts and what they learn about the codebase. The mechanism is plausible, and untested.
Self-hosted capacity: enterprise plumbing behind the beta
Cursor's enterprise story has three parts.
| Option | What it does |
|---|---|
| Self-Hosted Machines | Keeps tool execution inside your own network. Codebase, build artifacts and secrets stay on internal machines, while the agent handles tool calls locally. |
| My Machines | Connects a single laptop or VM to your account for personal workflows. |
| Team pools | Named worker queues for a team or company. Capacity grows as requests arrive and shrinks as workers disconnect, and idle machines sleep until a follow-up reconnects them. |
Pools are not tied to one repository: name a pool and any available worker can pick up the request. Cloud agents can also run on infrastructure a buyer already operates, including AWS Lambda, Coder, Cloudflare, Daytona, Modal, Namespace, Vercel and E2B. That list points at teams that will not hand build artifacts or secrets to a vendor's cloud, and whose security review stalls on exactly that point. The same pressure keeps pushing security checks earlier in the pipeline, from the release gate toward the keystroke.
What to watch: pricing, reliability, and the opt-out clause
Three gaps stand out. Cursor published no price for Projects, Origin, self-hosted pools or the infrastructure partners it named. That is a notable silence from a company already selling cheaper model routing as the answer to agent costs. Origin's rollout excludes Enterprise organizations whose admins choose not to participate, which suggests the company expects pushback from buyers it spent a section courting. And the autonomy claims, agents that follow a Slack channel or a pull request and act on what they find, depend on loops only a sustained deployment can test. Whether those loops hold in week three is not a question a launch post answers.
- Source : Cursor wants to run your agents and host the code they write — 2026-09-10
Get the tech essentials in 3 minutes every morning
One email, every weekday, with what actually matters in AI and tech.