Your multi-agent system made 400 tool calls, spawned six sub-agents, hit three external APIs, and executed a transaction. Six months later, a lawyer asks you to reconstruct exactly what happened and why. Your traces have a 30-day retention policy. Your prompts aren't versioned. Your model provider silently updated the endpoint. You can't answer.

I'm the person that lawyer hires. I've served as a court-appointed Special Master and expert witness in hundreds of federal and state proceedings, including In Re: Facebook, and I co-created the JAMS AI Dispute Resolution Rules. What I keep finding when I open up agentic systems is that the observability stack engineers already built for debugging is nearly, but not quite, the thing that would have saved them. The gap is usually four or five design decisions, each of which is cheap on day one and impossible to retrofit on day 400.
This is a talk about those decisions, told through real failed reconstructions. Not a compliance lecture. No legal background needed, and none will be assumed.
We'll cover: why span-level tracing is not the same as attribution, and what an agentic trace needs to carry to survive a challenge; the difference between logging what your agent did and logging what your agent knew; why prompt and model versioning is an evidentiary problem before it's an eval problem; the specific failure mode of RAG systems where retrieved context is gone by the time anyone asks what the model saw; and where non-determinism actually becomes a defense rather than a liability.