The Third Era Is Already Here

Three Eras in One Year

The speed of this transition is the real story.

Cursor just published "The third era of AI software development". Tab autocomplete was era one. Synchronous agents, where you prompt and watch the agent work in your editor, was era two. Now cloud agents, running autonomously on their own machines, are era three.

They put it plainly:

Cursor is no longer primarily about writing code. It is about helping developers build the factory that creates their software.

The numbers are wild. Agent usage in Cursor grew 15x in the past year. In March 2025, there were 2.5x more Tab users than agent users. Now that's flipped: 2x more agent users than Tab users. 35% of Cursor's own PRs come from agents running autonomously in cloud VMs.

The Tab era lasted about two years. The synchronous agent era may not last one.

I'm already past it.

Plan Locally, Build in the Cloud

Here's my workflow. I research and plan in Plan Mode locally. I think through the problem, break it down, write a clear plan. Then I ship it to a cloud agent and move on to other work.

Each PR the agent opens gets a Vercel preview. I review the work visually, in a real environment, not by reading diffs line by line. When something needs a fix, I @cursor in the GitHub PR comments. The same cloud agent picks it up.

What makes this safe: strong automated tests and proper CI/CD. Without those, you can't trust async agents. With them, most PRs just work. I rarely need to touch code locally.

For SOLAI, I shipped feature after feature this way in a very short period. A 3D editor with AI capabilities. Multiple complex features. All built by cloud agents while I focused on other things.

But the real turning point was long-running cloud agents. Cursor launched these with the ability to "plan first and finish more difficult work without human intervention." I built a full game with one. A single long-running agent researched the best tech stack based on my specs, built the entire game with complex systems and rules, verified the results on its own machine, and sent me a video demo. All while I wasn't watching. I didn't guide it. I didn't do follow-ups. It just worked.

From Pair Programming to Delegation

Nader Dabit nailed the framing: local agents are pair programming. Cloud agents are delegation.

Most engineers, even AI-powered ones, are still in pair-programming mode. They watch agents work in their editor. Do follow-ups by hand. Touch code. That was me a few months ago, babysitting local agents, doing manual fixes, feeling like Tab completion was the peak.

The actual shift is different. You stop writing code. You stop watching agents. They grind for hours, verify their own results, and come back with artifacts: videos, screenshots, live previews. Cursor's own team put it clearly: "the role of a developer is more about setting direction and deciding what ships."

This happened within a year. From heavy Tab usage to fully not touching code, mainly reviewing it. Sometimes not even reviewing it.

The Superpower

Here's what nobody talks about enough.

A customer tells me about a problem on a call. During a break, I spawn a few cloud agents to solve it. Minutes later, I send the customer a video demo of the fix. Built by an agent, not me. The customer saw the problem and the solution within the same hour.

I can build while talking to customers, while doing sports, while taking care of my loved ones. These agents don't sleep. They don't lose focus. They grind non-stop, follow your guidance, verify their results, and show you proof.

Even engineers who use AI daily aren't accepting what's already in their hands. These are superpowers. And they're here now.

It feels like a movie. But wallah, it's real.