Your Agent Didn’t Fail. Your Harness Did.

Official Schedule Context

AI agents do not fail only because the model is wrong. Many production failures happen in the

harness around the model: state is not persisted, two runs mutate the same session, a tool call

never returns, an approval loses scope, or an internal success never becomes user-visible proof.

This talk uses OpenClaw as a public case study to examine real harness failure modes and extract a

reusable production model for AI engineers. We will look at how events enter an agent system, how

session state is rehydrated, why single-writer lanes and throttles matter, and why tool execution

needs scoped approvals and auditable receipts. The core idea is simple: a model proposes, the

harness commits, and the receipt proves it. Attendees will leave with a practical 'run receipt'

audit they can apply to their own agents: what woke it up, which state did it inherit, what

authority did it use, what executed, and what evidence survived.

No official session recording transcript was found by exact title match on the AI Engineer YouTube channel during this run.

Pending transcript synthesis when an official recording or confirmed matching video is available.