Skip to content
TopicTracker
From HackerNewsView original
TranslationTranslation

Show HN: Retrace fork a failed AI agents run, replay it, prove the fix

Retrace is a tool that records AI agent runs, allowing users to replay steps, fork from any point to apply a fix, and share the result via a link.

Background

Retrace is a debugging tool purpose-built for AI agent workflows — the kind where a large language model (LLM) makes a sequence of tool calls (searching a database, calling an API, writing a file) to complete a task. Unlike traditional software, these runs are non-deterministic: the same prompt can produce different results, and failures are often buried in long log chains. Retrace records every step of a run so developers can replay it identically, "fork" (branch) from a specific step after making a code or prompt change, and share a stable link to prove the fix. It competes with or complements LangSmith, Langfuse, and other LLM observability platforms, but emphasizes reproducibility and collaborative debugging over mere monitoring. The tool is relevant to anyone building agents with frameworks like LangChain, CrewAI, or AutoGen, where tracing and replay remain ad-hoc pain points.

Related stories