Run ten coding agents at once with written briefs and a manager agent

Jackson Ly, working alone, ran Claude Code agents on a Mac and a home PC for 30 days, from 15 September to 14 October, often with ten or more running together and agents working through the night. At first a few agents at once was more than Ly could keep up with. Ly changed the work so it needed less attention: each agent got a written brief and a check for done fixed before it started, kept its state in files and a shared task list, and one manager agent on the PC tracked who was doing what and passed up only the decisions that needed a person. The agents worked across 21 repositories, rebuilding a SaaS product, building an event platform and a directory site, writing tax and legal explainers, and doing most of the QA by design parity checks and golden tests. Ly also ran Codex in goal mode on an open-source project, and later moved worker sessions on the PC to the cheaper GLM-5.3 model.

Average concurrency went from 2.0 agents to 6.9, and hours with an agent running from 38 to 145 a week. In 30 days Ly merged 1,029 pull requests (406 features, 293 fixes), about three agent-hours each, on 225.1 billion tokens, about $97,000 at list price, though the actual cost was flat plan fees. Almost 98% of tokens were cached context re-reads. One Codex goal run went 4.5 days with zero commits because its stop condition needed a step only Ly could do; after a rewrite the next session made 17 commits. Fresh reviewer agents still corrected 11 of 64, 11 of 44 and 17 of 26 statements in explainers the writers had marked as checked.

Most of the subagents named no model and so inherited Opus from their parent, even though Ly's rule was to run cheap models for execution. Set a model on every subagent. An overnight usage limit paused 30 agents for four hours; they resumed without loss only because state lived in files. The pull request count includes Ly's own work, and the post was written with AI help.