Breken is for engineering teams that already have tests and still learn about failures from users. In an October 2026 benchmark of 33 open-source products, 71% of the user-reported bugs only appeared when the product was running, with real data, another user, an outside failure, or time passing, and the median bug had sat in the code for 152 days. The launch is meant to run that check on the real application before a customer does. It rebuilds the application on its own machines and answers external services inside an isolated environment, so nothing is mocked or reachable from outside. History, including accounts, records, pending work, and unclosed sessions, grows from simulated use rather than fixtures. Synthetic users and agents, each with intent, habits, and limited patience, then work the product while Breken changes one condition at a time, such as a retry, a lost response, a clock jump, or a deploy in the middle of a job. A failure counts only if it reproduces every time against a control that never fails. The output is the exact state, a trace, and a draft pull request. Sarthak Ahuja and Arhan Singhal are building the company in San Francisco. Before the announcement, maintainers had merged Breken fixes for 205 bugs across 89 open-source projects, including PostHog, Supabase, Sentry, and Docker. Of 235 bugs already present on 1 July 2026 and reported by users afterward, Breken found 76, and 61 of those were proven by a reproduction that failed before the fix and passed after. The strongest code-reading model in the same comparison found 37, with 59 false alarms against Breken's 20.