Retrace Avis 

Écrit par l'entreprise

Retrace is an execution replay engine for AI agents. If you build with LLMs, you already know the hard part isn't writing the agent it's figuring out what went wrong when it misbehaves in production. Retrace exists to make that loop fast: find the exact step that broke, change one thing, and prove the fix. You add a single wrapper to your agent function a decorator in Python or a wrapper in TypeScript and from then on every LLM call, tool call, retrieval, and error is captured automatically as a "span" inside a "trace." Calls to OpenAI, Anthropic, and Gemini are auto-instrumented, so there's no manual logging to maintain. Open the dashboard and your run becomes an interactive timeline you can play, pause, and scrub through like a video. What makes Retrace different is forking. When you find the step where things went sideways, you can fork from that point, change the input, and the entire downstream agent re-executes not just one call. You get a side-by-side diff of the original run versus the new one, so you can actually see whether your change improved things, regressed, or did nothing. Beyond replay and forking, Retrace includes automatic failure detection (hallucinations, loops, goal drift, context loss, schema violations), circuit breakers and guardrails that can stop a runaway agent before the next expensive call, semantic search across all your runs, persistent agent memory, cost analytics, and CI regression gates. You can also publish any run as a shareable, interactive public link. Retrace works with Python and TypeScript SDKs, supports OpenTelemetry ingestion, and redacts PII on every ingestion path. There's a free tier to start, with paid plans as your usage grows.Voir plus