Retrace

Retrace enables debugging of AI agents through replay and forking. You can record an agent's entire execution, then replay it step by step to understand what happened. At any point, you can fork the run—creating a divergent path—to test alternative decisions without re-running the whole scenario from scratch. This makes it easier to reproduce flaky behaviors and compare different reasoning paths.

Its main advantage is making agent behavior transparent and explorable, much like a step-through debugger for traditional software. This is especially useful for complex, non‑deterministic agents where logs alone might not reveal the root cause. Forking supports “what‑if” analysis, helping to evaluate agent robustness.

However, available information is minimal—the sole evidence is a Product Hunt submission. We cannot confirm technical details, API/SDK availability, supported agent frameworks, stability, or performance overhead. The learning curve, deployment model (cloud, self‑hosted), and cost are unknown. Potential drawbacks such as incomplete state serialization or limited language support may exist.

It is likely suitable for engineers building or evaluating AI agents (e.g.