Quality assurance · Production

Cursor scales hundreds of concurrent autonomous coding agents to run for weeks on large software projects

The problem

Single coding agents handle focused tasks well but are too slow for complex projects that typically take human teams months to complete; running multiple agents in parallel creates hard coordination problems including lock contention, risk-averse behavior, and duplicated work.

First attempt

File-based locking for self-coordination created bottlenecks and brittleness; optimistic concurrency control was more robust but agents without hierarchy avoided hard tasks and made no real progress; an integrator role added for quality control created more bottlenecks than it solved.

Workflow diagram · grounded in source

Ambitious goal assigned

trigger

“To test this system, we pointed it at an ambitious goal: building a web browser from scratch.”

Planner agents create tasks

ai_action

“Planners continuously explore the codebase and create tasks. They can spawn sub-planners for specific areas, making planning itself parallel and recursive.”

Workers execute and push

ai_action

“Workers pick up tasks and focus entirely on completing them. They don't coordinate with other workers or worry about the big picture. They just grind on their assigned task until it's done, then push their changes.”

Judge validates cycle

validation

“At the end of each cycle, a judge agent determined whether to continue”

Fresh iteration restarts

feedback_loop

“the next iteration would start fresh”

Reported outcome

A planner/worker/judge hierarchy enabled hundreds of concurrent agents to run for weeks, producing over a million lines of code on ambitious projects including a from-scratch browser, a video rendering improvement of 25x that was merged to production, and a Solid-to-React migration passing CI.

Reported metrics

Lines of code written (browser project)over 1 million lines of code

Files created (browser project)1,000 files

Duration (browser project)close to a week

Video rendering speed improvement25x faster

Show all 14 reported metrics

lines of code written (browser project)over 1 million lines of code

files created (browser project)1,000 files

duration (browser project)close to a week

video rendering speed improvement25x faster

duration (Solid-to-React migration)over three weeks

edits (Solid-to-React migration)+266K/-193K edits

Java LSP project commits7.4K commits

Java LSP project lines of code550K LoC

Windows 7 emulator commits14.6K commits

Windows 7 emulator lines of code1.2M LoC

Excel project commits12K commits

Excel project lines of code1.6M LoC

tokens deployed across agentstrillions of tokens

concurrent worker agentsHundreds of workers run concurrently

Reported stack

GPT-5.2GPT-5.1-CodexOpus 4.5GitHubRustReactSolidCI

Source

https://cursor.com/blog/scaling-agents

Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

A planner/worker/judge hierarchy enabled hundreds of concurrent agents to run for weeks, producing over a million lines of code on ambitious projects including a from-scratch browser, a video rendering improvement of…

What tools did this team use?

GPT-5.2, GPT-5.1-Codex, Opus 4.5, GitHub, Rust, React, Solid, CI.

What results were reported?

Lines of code written (browser project): over 1 million lines of code; Files created (browser project): 1,000 files; Duration (browser project): close to a week; Video rendering speed improvement: 25x faster (source-reported, not independently verified).

What failed first in this deployment?

How is this quality assurance AI workflow structured?

Ambitious goal assigned → Planner agents create tasks → Workers execute and push → Judge validates cycle → Fresh iteration restarts.