Scaling long-running autonomous coding
simonwillison.net
January 20, 2026
3 min read
🔥🔥🔥🔥🔥
35/100
Summary
Wilson Lin at Cursor has experimented with a large fleet of autonomous coding agents, coordinating their work to generate over a million lines of code and trillions of tokens. The process involved running planners and sub-planners to create tasks for workers to execute.
Key Takeaways
- Cursor ran hundreds of autonomous coding agents for nearly a week to build a web browser from scratch, resulting in over 1 million lines of code across 1,000 files.
- The agents utilized planners and sub-planners to create tasks, with a judge agent determining project completion.
- The FastRender repository includes build instructions and uses Git submodules to access various web specifications, enhancing the agents' coding capabilities.
- Initial skepticism about the project was addressed after the inclusion of build instructions, leading to a functioning browser window capable of rendering web pages.
Community Sentiment
Positives
- AI significantly reduces the cost of remixing existing solutions, allowing for rapid exploration of uncharted territories in coding and development.
- The ability of LLMs to pattern match through codebases showcases impressive capabilities, enabling developers to leverage AI for more efficient coding.
Concerns
- LLMs often mimic intelligence rather than exhibit true understanding, leading to a reliance on human judgment to assess their outputs.
- Concerns about the environmental impact and costs associated with AI models, such as token and electricity costs, raise important questions about sustainability.
