Invited talk, University of Oxford, Oxford, UK. October 2026.
Agent systems fail quietly: a retrieval that returns the wrong memory, a cached context that is reused out of place, a tool result that is trusted without checking. This talk proposes three primitives that make such failures visible and controllable. Trust: deciding which inputs, tools, and memories an agent may act on, and at what confidence. Memory: evaluating memory systems jointly with the retrieval regime and backbone that use them, where rankings change with access strategy and instrumentation separates integration defects from capability gaps. Reuse: reusing context and computation across turns and sessions without reusing errors. I draw on MemoryCraft and ContextPilot and on agents deployed on edge devices and robots.
This is an executable talk in the edtrace format used by Stanford CS336: the Python program is the slide deck. Text lines are the narration, code lines run and can be stepped through, and inspected values appear inline.
Powered by Jekyll and Minimal Light theme.