Cursor announced the third era of software development with cloud AI agents

Cloud AI agents are becoming the primary form of software development: they perform complex tasks independently, work for hours on dedicated virtual machines, and require minimal human oversight. At Cursor, autonomous agents already create more than a third of all pull requests, and the company has noticed that developers adopting this approach become noticeably more productive.

From Tab to Agents

When Cursor first started, code was written manually, one character at a time. Tab autocomplete was the first game-changer—it identified routine tasks and automated them by offering context-based completions. The first era lasted nearly two years and delivered massive productivity gains for most developers.

Then models improved. They began working with larger contexts, using more tools, and executing longer chains of actions. Developers moved to synchronous agents—those that work in real-time in the editor and require constant supervision. The second era seemed more powerful, but it didn't last long.

The statistics show the transition clearly. In March 2025, there were 2.5 Tab users for every agent user. But within just a few months, the picture flipped—there are now twice as many agent users. Agent usage grew 15-fold over the year.

How Cloud Agents Work

Cloud agents work fundamentally differently. Instead of assisting a developer on a local machine in real-time, they receive a task and work independently on a virtual machine in the cloud—for hours if needed. They test the solution, iterate, run the full test suite, and verify integrations. Then they return logs, execution videos, and finished artifacts instead of a set of diffs.

This fundamentally changes the developer's role. They no longer guide the agent at every step or provide hints when errors occur. Their job is to define the task, set acceptance criteria, review the result, and provide feedback if revisions are needed. Most importantly, they can run multiple agents in parallel, since each operates on its own machine and doesn't compete for local resources.

Practice Inside Cursor

Within Cursor itself, developers are already working this way at scale. 35% of all pull requests they merge are created by cloud agents working completely independently. These aren't isolated cases—it's becoming standard practice. Developers have noticeably shifted their approach:

• Agents write nearly 100% of the code for the task
• Developers focus on decomposing large tasks into subtasks
• They review artifacts and execution videos, providing feedback when revisions are needed
• They run multiple agents in parallel instead of one

This resembles managing a team of junior developers more than using an autocomplete tool.

What This Means

The third era of development transforms the very nature of the programming profession. Instead of writing code, the programmer becomes an architect of tasks—describing the problem, setting criteria, and critiquing solution quality. This demands next-generation tools. Agents must be maximally independent, with full access to tools, complete project context, and stable execution environments. For companies building such tools, this means completely rethinking platform architecture.