---
title: "Scaling long-running autonomous coding"
description: "The post reports experiments running coding agents autonomously for weeks, with hundreds concurrently contributing to projects that typically take human teams months. It describes initial dynamic coordination through shared files, locks, and optimistic concurrency control, then explains why flat peer structures caused bottlenecks, fragility, risk aversion, and stalled progress. The successful design separates recursive planners from task-focused workers, uses a judge between cycles, and lets workers push changes to a shared branch with minimal conflicts. In tests, agents built a browser with more than 1 million lines across 1,000 files, migrated Solid to React over three weeks with +266K/-193K edits, and delivered a 25x faster Rust video renderer; other projects remain underway. The authors say GPT-5.2 models perform better on extended autonomous work, role-specific models and carefully tuned prompts matter, and moderate structure outperforms both flat peer coordination and elaborate controls, though drift and scheduling remain unsolved."
---

# Scaling long-running autonomous coding

[Cursor](https://yomu.fyi/company/cursor) · Wilson Lin · Jan 14, 2026

**Type:** Explainer

## Summary

The post reports experiments running coding agents autonomously for weeks, with hundreds concurrently contributing to projects that typically take human teams months. It describes initial dynamic coordination through shared files, locks, and optimistic concurrency control, then explains why flat peer structures caused bottlenecks, fragility, risk aversion, and stalled progress. The successful design separates recursive planners from task-focused workers, uses a judge between cycles, and lets workers push changes to a shared branch with minimal conflicts. In tests, agents built a browser with more than 1 million lines across 1,000 files, migrated Solid to React over three weeks with +266K/-193K edits, and delivered a 25x faster Rust video renderer; other projects remain underway. The authors say GPT-5.2 models perform better on extended autonomous work, role-specific models and carefully tuned prompts matter, and moderate structure outperforms both flat peer coordination and elaborate controls, though drift and scheduling remain unsolved.

## Context

The work aims to determine how far autonomous coding can scale for projects that typically require human teams months to complete. The experiments focus on coordinating many agents on one codebase over extended periods while avoiding conflicts, duplicated work, bottlenecks, drift, and loss of focus.

## Approach / What changed

The system evolved from shared-file coordination with locks and optimistic concurrency control to a role-based pipeline. Recursive planners create tasks, workers complete them independently and push changes to a shared branch, and a judge decides whether another cycle should begin. Model selection and prompting are tailored to different roles.

## Takeaways

- Lock-based coordination caused long waits, unreleased locks, and failures; twenty agents sometimes delivered the effective throughput of only two or three.
- A planner-worker hierarchy enabled hundreds of concurrent workers to contribute to a browser project exceeding 1 million lines across 1,000 files, with minimal conflicts on a shared branch.
- GPT-5.2 models performed better on extended autonomous work than the alternatives tested, while different models were assigned to planning and coding roles based on their strengths.

**Tags:** [AI Agents](https://yomu.fyi/topic/ai-agents), [Distributed Systems](https://yomu.fyi/topic/distributed-systems), [Scalability](https://yomu.fyi/topic/scalability)

- Source: [Cursor](https://cursor.com/blog/scaling-agents)
- Source URL: https://cursor.com/blog/scaling-agents
- Ingested by Yomu: 2026-09-02T11:44:50.832Z

[Read original post](https://cursor.com/blog/scaling-agents)