Building recursive agent systems

At Cursor， we run thousands of agents to help us train the next version of Composer.

We give them research tasks， and if they aren't succeeding or run into issues， they DM us on Slack or page us via PagerDuty.

Scaling training for Composer

We've built an org chart of agents that work together.

As we've scaled training for Composer， we've wanted to run thousands more experiments. This was possible before， but it was slow and hard to keep track of every experiment's status. To speed things up and parallelize work， we built an always-running agent system （yes， it's a loop）.

An agent system for research

Here's how the system works：

The main agent runs on a massive remote machine with all the tools you'd use locally， plus a file on disk acting as an "inbox" for the fleet.

It SSHes into machines running hundreds of child agents and collects their statuses into the inbox.

On every loop， it checks fleet health， keeps healthy tasks running in the background， and surfaces anything broken to the team on Slack.

Like all infra， the agents occasionally hit transient issues or need to be poked， so the main agent can control the whole fleet， quitting or restarting processes as needed.

This "fleet manager" builds on our previously published research on long-running agents. We've given the manager many different skills that encode tacit knowledge for how to run ML experiments， review and monitor results， and more.

Researchers with superpowers

Training a great model means trying a bunch of ideas for creating useful RL data.

A single laptop is not enough here， you really want an army of computers in the cloud to run experiments in parallel. And since we aren't compute-constrained， we rolled out this infra for everyone in ML.

Researcher time is our scarcest resource and we've found a way to scale their leverage by orders of magnitude. Imagine if you had a human manager with 10，000 direct reports. Obviously that wouldn't work well， but this human → agent "org" kind of does！

If you have a problem that is verifiable， where throwing more tokens at it will solve it faster or better， it's worth considering building a system like this. It's enabled us to have swarms of agents crawling through Composer's data to recursively improve itself for future versions.

And if this sounds exciting， we're hiring！

Lee Robinson@leerob · X

61导出 Markdown

2026-06-13 00:22·20天前

在 X 看原推· x.com

AI 摘要

Cursor 为训练下一代 Composer，构建了一个始终运行的递归智能体系统。主智能体在远程机器上通过 SSH 管理数百个子智能体，将状态收集到磁盘“收件箱”，循环检查集群健康并保持任务运行，通过 Slack 向团队报告问题。主智能体具备多种技能用于运行和监控 ML 实验。研究人员可并行运行数千个实验，大幅提升效率。对于可验证的问题，投入更多 tokens 能更快解决。

http://x.com/i/article/2065439304785039360

Building recursive agent systems

At Cursor， we run thousands of agents to help us train the next version of Composer.

We give them research tasks， and if they aren't succeeding or run into issues， they DM us on Slack or page us via PagerDuty.

Scaling training for Composer

We've built an org chart of agents that work together.