Agent Replay
Standalone forensic reconstruction for AI-agent executions. Rebuild what happened, identify the first provable divergence, and preserve the evidence supporting that conclusion.
What it does
Agent Replay is a standalone incident-reconstruction tool for AI-agent executions.
Agent systems can generate large amounts of telemetry, logs, traces, and tool-call records, but those records do not automatically explain what went wrong. Agent Replay converts execution evidence into a structured reconstruction that identifies what happened, how events relate to one another, and where the first provable divergence occurred.
Its central question is: What exactly happened, where did execution first diverge, and what evidence supports that conclusion?
Agent Replay can normalize execution records, validate temporal and provenance relationships, reconstruct causal sequences, compare expected and observed state, and produce machine-readable evidence explaining the divergence. Current capabilities include OpenTelemetry OTLP JSON ingestion, canonical event normalization, temporal validation, provenance checking, causal reconstruction, expected-versus-observed comparison, evidence hashing, machine-readable reconstruction output, and optional TRACE trust-record verification.
Agent Replay is intentionally standalone. DDC and other assurance adapters can provide additional evidence, but they are not required to use the core product. It is not an observability platform, monitoring service, policy engine, or agent runtime. Its purpose is forensic reconstruction after or during an incident.
Current public hardening line: v0.4.3.
Who it helps
AI-agent developers, platform engineers, security teams, incident responders, agent-runtime developers, MCP developers, auditors, and organizations operating autonomous systems.