Cursor's New AI Agents Just Changed Everything

Cursor's new AI agents just changed everything, so Cursor just changed how software gets built. Not gradually, not slowly, yesterday. If you're a business owner, an agency owner, a freelancer, anyone who pays people to build and maintain software, this one matters. Here's the number you need to hear first.

30%. That's how much of Cursor's own code is now being written by AI agents. Not humans, agents. Running on their own virtual computers, writing code, testing it, recording video proof of their work, then handing a merge ready pull request to a human to approve.

3 out of 10 code changes at Cursor are done by AI. That's not a case study, that's not a pitch. That's a company that built this tool using our on their own product right now in production. And yesterday, they just shipped something that removes the last major reason serious companies couldn't do the same thing.

Self-hosted cloud agents. And stay with me, because I'm going to break down so it's completely clear and easy to understand exactly what this is, why it matters, and why this is a bigger deal than most people are going to realize this week. Let me start with what a Cursor cloud agent is. It's not a autocomplete tool.

It's not a chatbot you type into and get your own code back. It is an AI that gets its own computer, its own virtual machine, its own terminal, its own browser, its own full desktop environment. You give it a task and it clones your code base, sets up the development environment, writes the code tests, tests it, and then fixes bugs and even records a screen video showing exactly what it built and how it works and then creates a pull request so that it's ready to merge. Now, you could come back, and that could be an hour later, or it could be the next morning, and watch a 30 second video of what the agent actually built.

You approve it, we'll give you feedback, and it just keeps going. That's the workflow. Agents get their own computer. Agents do the work for you, and agents show you exactly what they did and you decide what to do with it.

Do you want to publish it or not? And here's the thing that separates Cursor from every other AI coding tool right now. The agent doesn't just write the code and hope it works. It actually uses the software it built to test it.

It clicks buttons, it fills in forms, it navigates the app the way a real user would. Cursor's own data, sandboxed agents, the ones running in their own isolated environment, stop 40% less often than agents without one, meaning they actually finish tasks. They don't get stuck halfway and need babysitting. They complete the work and then hand it back.

That's a huge difference in practical usefulness. Now here's the problem that existed until yesterday. Cursor's cloud agents were already doing all of this, but they were running on Cursor's servers. Your code went to their cloud, your secrets went to their cloud, and you built artifacts that went to their cloud.

For a small startup, no problem. But think about who can't do that? Banks. Every financial institution who has compliance rules that say customer data, proprietary code, internal secrets, none of it can leave their controlled environment.

That's not a preference, that's a regulatory law. And healthcare companies, it's the same thing. You've got HIPAA compliance, you've got patient data, internal systems, nothing leaves the network. Government contractors too, defense companies, even legal firms, any company with client confidentiality obligations.

These companies wanted to use Cursor's agents. The agents are genuinely powerful, but the data had to leave the building to use them, and that was a non-starter. So what did some of these teams do? Well, they started building their own background agent systems from scratch.

Internal engineering resources diverted away from actual product work just to maintain agent infrastructure that gave them something close to what Cursor already had. That was the workaround, that was the cost. Cursor just made all of that unnecessary. Self-hosted cloud agents means the agent runs on your machines inside your network.

Your code never leaves, your secrets never leave, your build artifacts never leave. The agent is executing on your infrastructure with access to your internal tools, your cache, your private network endpoints, exactly the way one of your own engineers would work. Cursor handles the orchestration. So for example, the AI model, the planning, the user experience, you handle where the execution happens.

Here's how you actually set it up. So it's one command, agent worker start. That's it. That spins up a worker process on your machine that connects outbound to Cursor's cloud over HTTPS.

So there's no inbound ports to open, no firewall rules to change, no VPN tunnels to configure. The AI brain is in Cursor's cloud. The actual execution, every file it touches, every tool it runs, every line it writes, well that happens on your machine instead. And each agent session gets its own dedicated worker.

Workers can be long-lived or spin up for one task and tear down when it's done. If you need to scale to thousands of workers, Cursor actually gives you a Kubernetes operator and a helm chart. And you define the pool size and the system handles scaling updates and lifecycle automatically. If you're not on Kubernetes, there's actually fleet management now, which is an API that lets you build auto-scaling on any cloud infrastructure.

This is built to run at enterprise scale. For example, one financial services company with close to a thousand engineers said they'd been waiting for this. They'd already built a workflow so engineers can trigger a Cursor agent directly from Slack, have the agent to build the requested change and come back with a merge ready request. A thousand engineers, Slack messaging, pull request out, agent did the work in between and let that picture sit for a second.

Now, if you're watching this and thinking, I need to be using AI automation in my business, but I don't know where to start. Come join us in the AI Profit Boardroom at AIprofitboardroom.com. There are 2,600 business owners, agency owners and creators in there already using AI automation to get more done. Four weekly live coaching calls, daily tutorials with step-by-step walkthroughs, 30 day roadmaps, prompts for everything.

And a local map so you can connect with members near you and meet up in person. And there's always someone online, 24 seven. So if you need help, link in the comments description or just go to the AIprofitboardroom.com. Notion said this publicly, running agent workloads in their own cloud environment lets agents access more tools more securely and saves a team from needing to maintain multiple stacks.

Before this, the Notion was maintaining their own internal agent infrastructure just to get around the security problem. They were building the work around themselves. Now that's gone. Cursor handles the orchestration.

Notion runs it inside the walls. Everyone moves faster. And these agents could be triggered from anywhere, from the cursor editor, from the cursor web app, from Slack, from GitHub, from Linear, from Webhook, on a schedule, on an event, from the API. An agency managing client accounts could set up automation so that every bug ticket that hits a certain priority in Linear automatically spins up a cursor agent to investigate, to file and affix and come back with a pull request.

The account manager watches the video, approves, affix, done. No developer needed for that loop. The human in the quality gate, not the labor. Let me give you the broader picture of who's already in on this.

Salesforce. Over 90% of their developers out of 20,000 developers now use cursor. They saw double digit improvements in how fast code ships and how fast pull requests move through review. A senior Salesforce executive publicly called cursor his favorite enterprise AI service and said every one of their 40,000 engineers is now assisted by AI with productivity going up incredibly.

NVIDIA uses cursor too. PwC uses them. Stripe, over half of the Fortune 500 actually uses cursor. NotPilot's production deployments with measurable results.

And the companies that couldn't join them, for example because of security requirements, the banks, the healthcare systems, the defense contractors, well that door is now open with the cloud AI agents. So here's what actually happens in one of these agent sessions. Completely concrete, cursor's own team needed to replace a static label in the app with one that dynamically showed the number of linter errors in file. They gave the task to a cloud agent.

The agent implemented the feature with styling to match the existing CSS, tested two real cases inside the cursor desktop app, a file with multiple type errors and a clean file with no errors, verified both worked correctly and then recorded the whole session on video, rebased onto main, resolved merge conflicts and squashed to a single commit. The human watched a video, confirmed it works, approved, done. That task from brief to merge ready PR with zero human involvement in the actual execution. For an agency owner, you could for example brief an agent on a client change request in the morning.

Then you could come back after your next call to a screen recording of the completed work, ready for your review. Watch 30 seconds, approve or give a no and that's your development review process done now. And let me go back to the video artifacts for a second. So when a cursor agent finishes, it records itself.

You get a real screen recording, the agent navigating the app, clicking buttons, testing the feature, showing you exactly what it built and how it behaves. Watching a 30 second video is far faster than 500 lines of code checking, right? So it's way faster and it tells you more. You can see if the UI looks right.

You can see if the form works. You can see bugs that would pass every automated test because the automated test wasn't checking the visual. And this is what makes the handoff actually work in process. So it's not like blind trust in the AI agent.

It's a clear artifact that shows you exactly what happened in this video and lets you make a fast, confident decision. Self-hosted agents keep all of this same isolation, same video artifacts, same quality, just running inside your network instead of on cursor servers. So you lose nothing, but you gain complete data sovereignty. And here's what I keep coming back to.

The teams winning right now are not the ones with the most people. They're the ones where every person is directing the most agents, right? One person, 10 agents running in parallel, each one working through its task whilst the human focuses on decisions, client strategy. That's a new unit of output.

And the speed of the shift is the part that's hard to hold in your head. We went from AI that suggests the next line of code to AI that gets his own computer, builds a whole feature, records a proof, and then hands it back a pull request in about 18 months, right? 30% of Cursor's code is agent created today. What's that number going to be in 12 months?

Cursor's own state division is what they call self-driving code bases. Agents that don't just write code on request, but agents that merge pull requests, manage rollouts, monitor production. The entire software development loop with humans as the decision layer, not in the execution layer. That future moved closer yesterday, not because of one feature, because the last major objection, security, just got answered.

The companies moving now will have the structural advantage when the rest of the industry catches up. Every agent created pull request frees a human up for the harder problems. Every automated workflow is a task that never lands on a person's plate. The window is open right now.

And if you want to be in that group that's moving, come join the AR Profit Boardroom. Link in the comments description or go to the arprofitboardroom.com. 2,600 business owners, agency owners, creators, solopreneurs, already using AI automation to work at a completely different level. With four weekly live coaching calls, daily tutorials, 30-day roadmaps, and prompts for everything.

Plus a local map to find and meet members near you. And someone is always online 24-7, so whenever you need help, link is in the description. The wall came down yesterday, now it's about speed.