How the Creator of Claude Code Actually Uses It: Thousands of Agents Running Overnight

Boris Cherny, who built Claude Code at Anthropic, revealed his actual workflow at a recent event -- and it is far beyond what most users imagine. Five to ten sessions running simultaneously, hundreds of agents working in parallel, and thousands of sub-agents doing deeper work overnight. This is not theoretical; it is how the tool's own creator operates day to day.

The Scale of Operation

When asked how many sessions he typically runs, Cherny's answer was startling: five to ten sessions, each spawning a bunch of agents, totaling a few hundred agents running at any given time. Every night, a few thousand sub-agents tackle deeper work while he sleeps.

The Loop Pattern

The pattern Cherny has been using more and more is what he calls Slash Loop. The concept is simple: Claude uses cron to schedule a repeating job at some interval -- every minute, every five minutes, every day. He has dozens of these loops running for different purposes. One babysits his PRs, fixing CI issues and auto-rebasing. Another keeps CI healthy by catching flaky tests. The beauty is in the simplicity: it is just scheduled Claude Code sessions that run autonomously.

Managing at Scale

There are a few ways to manage work at this scale. One approach is asking Claude to use a bunch of sub-agents to distribute work across a problem. But the loop pattern has become his go-to because it is the simplest thing that works -- just cron jobs triggering Claude Code sessions that handle ongoing maintenance tasks without human intervention.

Key Takeaways

• The creator of Claude Code runs 5-10 sessions simultaneously with hundreds of agents and thousands of nightly sub-agents
• The /loop pattern uses cron to schedule repeating Claude Code sessions for autonomous maintenance
• Practical loop uses include babysitting PRs, fixing CI, auto-rebasing, and catching flaky tests
• The most effective patterns tend to be the simplest -- cron-scheduled sessions beat complex orchestration
• Sub-agents are best suited for deeper, overnight work that does not require immediate feedback

Resources

• Claude Code -- Anthropic's CLI tool for working with Claude
• Boris Cherny on Anthropic -- Creator of Claude Code at Anthropic