agentfork

agentfork is a developer tool that helps independent developers debug and improve large language model (LLM) agents by recording each step, replaying cached results to save costs, and running multiple variants of specific steps repeatedly to measure pass rates. It runs fully locally, requires no user accounts, and supports OpenAI and Anthropic models. agentfork enables efficient testing of prompt and model changes for LLM agents, helping pinpoint and verify fixes with evidence rather than anecdote.