I built an open source SDK to catch AI agent regressions before they ship.

Short summary

Replayd is an open-source Python SDK that captures failed AI agent runs as regression tests, then replays them before deployment to catch recurrence early. It handles LLM non-determinism by checking whether specific failures return rather than matching exact outputs—structural failures use deterministic assertions, while semantic failures get LLM judgment. Framework-agnostic and dependency-free in the core.

• •Open-source SDK (replayd) converts AI agent failures into regression tests
• •Handles non-determinism by checking failure recurrence, not output matching
• •Framework-agnostic, zero core dependencies, inviting production feedback

Generated with AI, which can make mistakes.