AgentReplay — time-travel debugger for AI agents

When an AI agent does something wrong, "read the logs" isn't debugging — you get a wall of text and no way to ask "what if this one step had gone differently?" And re-running the agent just produces a different conversation; the one that actually broke is gone. AgentReplay treats agent runs like flight data. A Python SDK records every LLM call, tool call, and state change as the agent runs — two-line integration, agent logic untouched. The dashboard lays the run out as a timeline. AI root-cause analysis points at the exact step that broke. Then you fork: replay from that step with a fix applied — live inference, temperature 0, tools sandboxed — into a new run that carries a parent_run_id, while the original stays immutable. A side-by-side compare shows the broken run against the fixed one, on the same conversation, with no real-world side effects. The demo is a real failure mode from my production lead-qualification agent, Nestaro: it booked a caller for Friday when they asked for Saturday. That run is replayed through the recorder — AgentReplay flags step 2 as the culprit and forks a corrected replay that books Saturday, proven side by side, without touching the real booking system. Stack: Python SDK · FastAPI + SQLModel + NeonDB (JSONB event store) · Next.js dashboard. Hosted analysis runs on OpenRouter, provider-swappable via env vars. AMD: AgentReplay's ROCm/PyTorch stack ran on an AMD Radeon gfx1100 (RDNA3, 48 GB), and its real, unmodified root-cause prompt was executed through Gemma 3 4B resident on the AMD GPU — full evidence committed in the repo's amd/ folder.