Show HN: Agentfork – Replay an LLM agent from any step and test fixes (github.com)

Agentfork, a new tool for LLM agents, allows developers to identify and fix issues more efficiently by enabling them to "fork" their agent at any point in its processing. Instead of rerunning the entire agent every time an issue crops up, users can replay previous steps from a cache and test multiple variations of a specific problematic step in parallel, significantly reducing the time and token costs associated with debugging. For instance, when faced with incorrect refund calculations, Agentfork can sweep through different model iterations and prompt adjustments to pinpoint effective fixes, offering concrete pass rates rather than anecdotal success. This innovation is particularly significant for the AI/ML community as it streamlines the debugging process, making it far less reliant on guesswork. By coupling traditional LLM capabilities with advanced testing features, Agentfork allows for a more methodical approach to optimizing agent performance. It records all model calls and tool interactions, runs local computations without requiring external data or accounts, and integrates seamlessly with existing frameworks like LangChain. This capability paves the way for more robust and reliable AI applications, enhancing developers' ability to iterate and refine their models systematically.