# Boris Cherny says he now runs thousands of Claude Code agents at once > Source: > Published: 2026-07-17 19:51:40+00:00 *Boris Cherny hasn't written a line of code by hand in eight months. He runs thousands of Claude Code agents overnight instead, and Anthropic is betting the rest of the industry follows.* Cherny built Claude Code. Now he barely touches it. At a Fortune Brainstorm Tech talk in June, he told Fortune's Jeremy Kahn that he hasn't written a line of code by hand in roughly eight months, and that Claude Code itself has been, in his words, "100 percent written by Claude Code" for over six months. The line traveled fast. A video clip of the talk pulled more than 870 posts on X within eleven hours, according to Fortune's reporting. What replaced hand-coding, and even hand-prompting, is something Cherny and others at Anthropic are now calling loop engineering. He doesn't write prompts turn by turn anymore. He writes loops: automated workflows that trigger on a schedule or an event, prompt Claude, evaluate what comes back, act on it, and start again. Some of his own loops have names, like babysit-prs, post-merge-sweeper, and pr-pruner, running on five, thirty, and sixty minute intervals to keep pull requests moving without him at the keyboard. The scale is the part that got attention. Cherny runs five terminal instances of Claude Code at once, plus five to ten more sessions through claude.ai/code, and says nightly orchestration regularly pushes into the hundreds or thousands of agents running for five to twenty hours at a stretch. On his most extreme days, he told Fortune, he is managing tens of thousands of agents at once. That is not a hypothetical roadmap slide. It is a description of his actual Tuesday. Cherny has been explicit that this isn't magic, and it isn't a single new tool either. He's described five conditions that make it work. Auto mode handles permissions, because an agent that waits for a click to proceed might as well not run overnight at all. Dynamic workflows let Claude write its own orchestration plan once you mention the word workflow in a prompt, so a hundred agents follow the same stages in order. Commands like slash goal or slash loop force the run to actually finish, since Cherny says agents left unchecked tend to complete around 60 percent of a task and declare victory anyway. Cloud execution through Anthropic's Routines lets him close his laptop while the work continues. None of it means anything without end to end self verification. If Claude can't check its own output against a real environment, an overnight run is a gamble, not a workflow. Anthropic says the internal numbers back it up. Cherny told Platformer that engineers inside the company are shipping roughly 70 percent more per head since loop driven work became the default. Take that with a grain of salt. It's a company reporting on itself. But it's a specific number, not a vague productivity boast. ## The competitive backdrop This lands the same week Moonshot AI released Kimi K3, a 2.8 trillion parameter open weight model that topped the Frontend Code Arena leaderboard at 1,679 points in blind developer testing, ahead of Claude Fable 5, according to Tom's Hardware and CNBC's coverage of the July 16 launch. Kimi K3 still trails Claude Fable 5 and GPT-5.6 Sol on Artificial Analysis's broader intelligence index, and Moonshot's own benchmark table mixes coding harnesses, so a clean model to model comparison isn't quite ready yet. Full weights are promised by July 27. Here's the thing. Kimi K3's pitch is a better single model. Cherny's pitch is a different shape of competition entirely. Anthropic isn't just racing to win the next benchmark table. It's arguing that the real edge sits in how many agents you can run at once, how well they verify their own work, and how little a human has to intervene while they do it. If that argument holds, raw model quality becomes one input among several, next to orchestration tooling that only Anthropic currently ships as a first party product. For engineering leaders, the immediate question isn't whether to buy in on day one. It's whether your team's review and verification infrastructure can survive agents working at this volume before you hand them the keys overnight. Cherny's own framework puts self verification last on his list of five rules. He's clear it's the one that decides whether a thousand agent run is progress or an expensive mess to clean up in the morning. **Also read:** [Meta Is Reportedly Negotiating to Rent Its AI Chips to Rival Anthropic](https://startupfortune.com/meta-is-reportedly-negotiating-to-rent-its-ai-chips-to-rival-anthropic/) • [ASML Gives Nearly Every Worker 20,000 Euros in Stock to Keep Them Through 2030](https://startupfortune.com/asml-gives-nearly-every-worker-20000-euros-in-stock-to-keep-them-through-2030/) • [Kimi K3 Shows the Cheap Chinese AI Era Is Coming to an End](https://startupfortune.com/kimi-k3-shows-the-cheap-chinese-ai-era-is-coming-to-an-end/)