Retrace

What it does

Retrace is an execution replay engine for AI agents. It records every LLM call, tool invocation, and error an agent makes, then lets you replay the run, fork from the exact step that broke, modify inputs, and re-run to verify a fix. The platform also provides runtime guardrails (cost budgets, loop detection), CI/CD eval gates, and failure classification (MAST taxonomy).

Who it is for

Retrace is built for developers and teams shipping AI agents in production. It targets AI engineers using frameworks like LangChain, CrewAI, LlamaIndex, or AutoGen, and supports Python and TypeScript SDKs with auto-instrumentation for OpenAI, Anthropic, and Google Gemini.

Why it matters

AI agent failures are often multi-step and hard to debug. Traditional observability shows what broke but not how to fix it. Retrace closes the loop: a production failure becomes a permanent regression test that can be re-run, forked, and verified. This reduces debugging time, prevents regressions from reaching users, and helps control costs with runtime guardrails.

Launch signal

Retrace is live with a free tier (1,000 traces/month) and paid plans starting at $29/month. The website shows a real recorded trace and a working product demo. The founder is Yashwanth. No funding stage or launch date is mentioned in the available material.

Brand and naming

The name "Retrace" directly evokes the core action of replaying and re-running agent traces. It is short, memorable, and domain-relevant. The tagline "Git branching for AI agent execution" positions the product as a familiar version-control metaphor for debugging, which helps developers quickly grasp the value proposition.