I'm Ayush, cofounder of Culprit, based in Bangalore. Culprit is git bisect for agent behaviour. When your agent gets worse, it replays your evals across your git history, uses repeated runs and 95% confidence intervals instead of single runs, and hands back the commit, the lines and a verified fix. It ships as a pip-installable CLI, with a CI check that gates pull requests. I build with Claude, LangGraph, Langfuse, Claude Code and Codex, and deploy on Vercel.

What's working: in our case study, Culprit named 0 wrong commits in 60 bisects, across 3 agents and 6 models. What I'm still figuring out is how teams actually fit regression checks into their eval habits, since agent evals are noisy and many suites have blind spots.

What I'd love from this community: an intro to a team shipping an agent in production who'd try Culprit on their repo and tell me honestly where it falls short.