Cursor recently experimented with using hundreds of AI agents to build a web browser; they ran for close to a week, writing 1M+ lines of code across 1,000 files
Scaling long-running autonomous coding. Wilson Lin at Cursor has been doing some experiments to see how far you can push a large fleet of “autonomous” coding agents:
Simon Willison's WeblogSimon Willison
Context & Ripple Effects
Cursor’s experiment makes orchestration—not just code generation—the key product question: a large agent fleet needs work to be divided, integrated, and checked before its output is useful. It follows Cursor’s earlier web and Slack controls for managing coding agents across devices, extending the company’s management layer from access to sustained execution.
Cursor gains a practical stress test for coordinating long-running autonomous agents across a large codebase, exposing whether integration, testing, and review can keep pace with generated output.
For developers evaluating Cursor, human oversight becomes the immediate constraint: high-volume agent output still has to be validated before it can be trusted in advanced projects.
Second-order effects
Competing coding-agent vendors face pressure to show not only capable individual agents but also reliable multi-agent coordination, a direction reflected in Cursor’s subsequent specialized autonomous coding-agent launch.
Demand shifts toward agent-management controls—assignment, monitoring, and review—rather than a chat-style coding interface alone, reinforcing the value of the web and Slack management layer.
Third-order effects
If large agent fleets become reliable, coding tools may compete increasingly on orchestration, verification, and the infrastructure needed to run persistent agents, rather than on model access alone.
The experiment also highlights a limiting condition: concerns about fragile “vibe-coded” foundations could make reliability and auditability the durable differentiators for autonomous software development.
The trend: AI coding is moving from single-assistant generation toward managed, long-running fleets of agents, with coordination and quality control becoming the central product challenge.
We built a browser with GPT-5.2 in Cursor. It ran uninterrupted for one week. It's 3M+ lines of code across thousands of files. The rendering engine is from-scratch in Rust with HTML parsing, CSS cascade, layout, text shaping, paint, and a custom JS VM. It *kind of* works! I…
@patrickc @cursor_ai Lots more to figure out, but perhaps a glimpse at the not-too-distant future. One concrete data point: a critical, isolated component of an upcoming launch was fully written by one of these “agent grind” sessions. It finished overnight and would've taken roug…
Claude Code and Opus were taking all the attention, so the Cursor team built a browser from scratch using GPT 5.2. 3M plus lines of code across thousands of files. Crazy.
Holy S***: CEO of Cursor said they coordinated hundreds of GPT-5.2 agents to autonomously build a browser from scratch in 1 week “We built a browser with GPT-5.2 in Cursor. It ran uninterrupted for one week. It's 3M+ lines of code across thousands of files. The rendering engine […
Cursor is becoming obsolete because Claude Code just does everything faster and better, so they are trying to hype “generative AI” when it's already looking lame. If you ask an agentic coding system with a runtime feedback loop to build an app whose meaning is clear, like “web
This is unbelievable! GPT-5.2 built a browser with 3 million lines of code in one week. It's practically game over for software development, as these models will get exponentially better this year and beyond. Incredibly, first level AGI for coding was reached in 2025!
Many people think it was 1 agent writing 3M+ lines of code. It's not. It was hundreds of concurrent agents. Key learnings from Cursor's blog: - Letting many agents self-coordinate as peers does not work - Clear roles work better: planners, workers, judges - GPT-5.2 performs be…
In case anyone's interested, that vibe-coded(ish) web browser project that Cursor released the other day does actually compile now, they fixed it up and added instructions to the README. I built it on my Mac and took these screenshots: [image]
Long predictions are falling faster all the time, but it's pretty notably that Simon only made this forecast on Jan 8, predicting something that might happen 3 years in the future. [embedded post]
Having compiled and run the web browser that Cursor built in a couple of weeks using mostly a giant fleet of coding agents I'm actually very impressed by it - there are rendering glitches but the renders it produces are surprisingly usable for a few-week-old project simonwillison…