Retrace

When an agent fails in production, the standard post-mortem is log archaeology: you scroll timestamps, guess at which LLM call returned the bad output, and ship a fix you cannot verify without re-deploying into live traffic. Retrace replaces that loop with a record-replay workflow. A single decorator instruments the agent, capturing every LLM call, tool call, and decision point as a structured trace. When a run fails, you fork from the exact failing step, edit the prompt or tool input, re-run the fork, and receive a verdict — fixed or not fixed — before the change ships.

The differentiating capability is the prove-the-fix verdict tied to CI/CD gating. Most observability tools tell you that something went wrong. Retrace lets you attach the recorded failure as a regression test, run your fix against it, and block the deploy if the fix does not hold. The vendor also describes a failure taxonomy called MAST, which auto-classifies why a run failed — groundedness gaps, drift, loop detection — so the diagnosis surfaces before a user reports a problem.

Retrace fits teams building agents in Python or JavaScript on any LLM provider — the vendor states compatibility with OpenAI, Anthropic, Gemini, and any framework via the SDK. The runtime guardrail layer, including budget limits, loop circuit breakers, and a pre-call gateway with hold-for-approval, addresses the budget blow-out risk that most tracing tools ignore entirely. The platform does not offer a self-hosted deployment path, which is the condition under which regulated teams or those with strict data residency requirements will route around it. Free usage is capped at 1,000 traces per month, after which continued access requires a paid tier.