About this session

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.

Topics

Speaker

Vinoth Govindarajan

Member of Technical Staff · Open AI