When an AI agent makes a bad trade, denies a customer, exposes sensitive information, or calls the wrong tool, “the model did it” is not a root-cause analysis. This talk shows AI engineers how to build a forensic flight recorder that can reconstruct exactly what a production agent saw, decided, and did.

Using a failed financial-services agent workflow as a running example, Daniel Garrie will break down an evidence-ready architecture that captures:

Correlation IDs across model calls, retrieval systems, tools, policy checks, and human approvals
Prompt, context, model, tool, and policy versions
Tool inputs, outputs, permissions, exceptions, and side effects
Integrity controls that reveal whether logs or records were altered.
Replay mechanisms—and the limits of deterministic replay
Retention and redaction controls for sensitive customer and business data
Evaluation hooks that distinguish hallucination, retrieval failure, tool misuse, and policy bypass

The goal is not to log everything forever. It is to preserve the minimum reliable evidence needed to debug failures, evaluate agent behavior, investigate incidents, and explain consequential automated actions.

Attendees will leave with a practical agent-event schema, a seven-part logging checklist, and an incident-reconstruction workflow they can adapt to production systems.