Agents are quickly shaping up to be the next big thing in coding.
In a post shared this week, Cursor CEO Michael Truell laid out a sweeping vision for how AI is reshaping software development — and backed it up with a striking internal data point: more than one-third of the pull requests merged at Cursor are now created by autonomous agents running independently in cloud virtual machines.
Truell framed the current moment as the dawn of a “third era” of AI-assisted development. The first was defined by tab autocomplete, which automated low-entropy, repetitive coding tasks. The second came with synchronous agents — AI systems that developers directed through back-and-forth prompt-and-response loops. Now, Truell argues, a third era is emerging, one defined by agents that can tackle larger, longer-horizon tasks with minimal human supervision.
“Cursor is no longer primarily about writing code,” Truell wrote. “It is about helping developers build the factory that creates their software.” That factory, in his telling, is made up of fleets of cloud-based agents that developers interact with more like teammates than tools — handing off tasks, setting review criteria, and evaluating outputs rather than writing line-by-line.
The numbers Truell shared suggest the shift is already well underway. As recently as March 2025, Cursor had roughly 2.5 times as many Tab users as agent users. That ratio has now inverted: agent users outnumber Tab users two-to-one, and overall agent usage on the platform has grown more than 15 times over the past year. Truell noted that many Cursor users today never touch the tab key at all.
The 35% internal PR figure is the most concrete signal yet of how far this transition has progressed. According to Truell, developers at Cursor who have adopted this new workflow share three defining characteristics: they have agents write nearly 100% of their code, they spend their time breaking down problems and reviewing agent-produced artifacts, and they run multiple agents simultaneously rather than guiding a single one to completion.
The mechanics enabling this shift are worth understanding. Unlike synchronous agents that compete for resources on a developer’s local machine and require constant oversight, cloud agents each run on their own virtual machine. A developer can hand off a task and move on, while the agent works through it over hours — iterating, testing, and ultimately returning not just a diff but reviewable artifacts: logs, video recordings, and live previews. That richer output is what makes parallel agent workflows practical at scale.
Truell is candid that significant work remains before this approach becomes standard across the industry. At scale, problems that a single developer can easily work around — a flaky test, a broken environment — can interrupt every agent run. Ensuring agents have reliable access to the tools and context they need remains an open engineering challenge.
Cursor, for its part, has been somewhat overshadowed by CLI platforms like Claude Code and Codex. But it now seems to be hoping to skip that part entirely, and move straight to agentic coding. The pace of change is striking. Truell suggested the synchronous agent era may last less than a year before being eclipsed by cloud-native, autonomous workflows. “A year from now,” he wrote, “we think the vast majority of development work will be done by these kinds of agents.” For software teams watching the landscape shift, Cursor’s internal numbers suggest that future may arrive sooner than expected.