Seeing Inside the Mind of AI: Tracing Agents, Tools, and Weird Decisions
AI is starting to feel less like software and more like a strange coworker with root access. It can read, write, reason, call tools, trigger workflows, move data, spend money, and make decisions across systems most people barely understand. We see the prompt. We see the response. But the interesting part happens in the middle, where AI-driven systems fan out through APIs, containers, queues, databases, serverless functions, model calls, SaaS platforms, and external services.
This talk is about opening up that middle. Michael Barbine will take the audience on a practical, funny, and deeply suspicious tour of what it means to instrument AI systems so we can see what they are actually doing. His approach comes from security, observability, automation, and an unusually disciplined obsession with games, rules, feedback loops, and measurable behavior. Whether debugging infrastructure, testing endpoint defenses, building agentic workflows, or playing tens of thousands of rounds of Rock Paper Scissors, the core question is the same: what happened, what changed, what pattern did we miss, and how do we prove it? Using distributed tracing, structured events, correlation IDs, audit trails, and purpose-built observability patterns, we can follow AI workflows across modern infrastructure and catch the moments where things get weird. Where did the agent hesitate? What did it call? What did it retry? What did it hallucinate? What did it spend? What data did it touch? What did it decide quietly? And why should anyone trust a system they cannot inspect?
This is a talk for hackers, builders, defenders, operators, and anyone who has ever looked at a black box and taken it personally. It connects classic hacker curiosity with one of the most urgent problems in modern computing: how to understand systems that are becoming more autonomous, persuasive, and embedded in everyday life. Attendees will leave with practical patterns for tracing AI applications, a clearer mental model for debugging agentic systems, and a renewed appreciation for the ancient hacker principle that mystery is not an acceptable interface.
Friday 1500 Crystal Ballroom