01What Cursor Self-Hosted Cloud Agents Are

Cursor self-hosted cloud agents are AI agents that run on your own machines inside your network, not on Cursor's servers. They get their own virtual computer, terminal, browser, and full desktop environment. You give them a task, and they clone your code base, set up the development environment, write code, test it, fix bugs, and record a video of what they built.

Then they create a pull request ready for a human to review and approve. The agent does the work. You decide what happens with it.

This is not an autocomplete tool or a chatbot. It is a full agent that acts like a remote engineer.

02Top Features of Cursor Cloud Agents

Cursor cloud agents come with several powerful features that set them apart. They run on their own virtual machines with a terminal and browser. They can use your internal tools, cache, and private network endpoints.

Each agent session gets its own dedicated worker.

  • Self-hosted execution: Code, secrets, and build artifacts never leave your network.
  • Video proof: Agents record their screen while testing, so you see exactly what happened.
  • Merge-ready pull requests: The agent rebases onto main, resolves conflicts, and squashes commits.
  • Trigger from anywhere: Start agents from the Cursor editor, web app, Slack, GitHub, Linear, webhooks, or the API.
  • Scale to thousands: Use the Kubernetes operator and helm chart, or the fleet management API.

These features make Cursor cloud agents practical for real development work, not just prototypes.

03How to Set Up Self-Hosted Agents

Setting up Cursor self-hosted agents is simple. You run one command and the worker starts. Here are the steps shown in the episode:

  1. Run the command agent worker start on your machine.
  2. The worker connects outbound to Cursor's cloud over HTTPS.
  3. No inbound ports, firewall rules, or VPN tunnels are needed.
  4. The AI brain stays in Cursor's cloud, but execution happens on your machine.
  5. For scaling, use the Kubernetes operator and helm chart to define pool size.
  6. Or use the fleet management API to build auto-scaling on any cloud infrastructure.

Workers can be long-lived or spin up for one task and then tear down. The system handles scaling, updates, and lifecycle automatically when you use Kubernetes.

04How to Use Cursor Agents in Your Workflow

Using Cursor self-hosted cloud agents fits right into your existing workflow. You give the agent a task, like replacing a static label with a dynamic one. The agent implements the feature, matches the CSS, and tests real cases in the actual app.

In the episode, Cursor's own team used an agent to fix a linter error counter. The agent tested two cases: a file with multiple type errors and a clean file. It verified both worked and recorded the session on video.

Then it rebased onto main, resolved conflicts, and squashed to one commit.

  • From Slack: A financial services company with 1,000 engineers triggers agents directly from Slack.
  • From Linear: Set automation so every high-priority bug ticket spins up an agent to investigate and fix.
  • From a schedule or event: Agents can run on a timer or when something changes.

The human watches a 30-second video, approves or gives feedback, and the agent keeps going. You are the quality gate, not the labor.

05Results and Comparisons From the Episode

Cursor's own data shows that sandboxed agents stop 40% less often than agents without one. They finish tasks instead of getting stuck halfway. And 30% of Cursor's own code is now written by AI agents in production.

Other companies see results too. Salesforce has over 90% of its 20,000 developers using Cursor, with double-digit improvements in code shipping speed and pull request review. A senior Salesforce executive said all 40,000 engineers are AI-assisted, and productivity is up incredibly.

Notion uses Cursor to save a team from maintaining multiple stacks. Half of the Fortune 500 uses Cursor, including NVIDIA and PwC.

Before this feature, banks, healthcare, and defense contractors had to build their own agent systems from scratch. That wasted engineering time. Now they can use Cursor directly.

06Tips, Limits, and Mistakes to Avoid

Self-hosted cloud agents are powerful, but you need to use them right. The episode gives a few clear warnings. First, always watch the video artifact before approving.

It is faster than reading code and shows visual bugs that automated tests miss.

  • Do not skip review: The human is the quality gate, not the execution layer.
  • Use sandboxed agents: They stop 40% less often, so let them finish tasks.
  • Plan for scale: Use Kubernetes or fleet management for thousands of workers.
  • Keep secrets safe: Self-hosted means code and secrets stay inside your network.

A common mistake is treating agents like autocomplete. They are not. Give them full tasks with clear goals.

Also, do not expect them to work without review. The video proof replaces manual code reading, but you still approve the merge.

07Who Should Use Cursor Self-Hosted Agents

This feature is for any company that builds or maintains software but has strict security rules. Banks, healthcare systems, government contractors, defense companies, and legal firms all benefit. They could not send code to Cursor's cloud before.

Now they can run agents inside their network.

It is also for agencies and freelancers who want to multiply their output. One person can direct 10 agents in parallel while focusing on client strategy and decisions. The episode calls this the new unit of output.

If you run a business that depends on software, this matters. The teams winning are not the ones with the most people. They are the ones where every person directs the most agents.

The security objection is gone, so the door is open for everyone.