Dan Farrelly divides agent architecture into execution, context and compute. Execution covers durable state, retries and control flow; context covers models, prompts, tools and memory; compute covers the environments where work runs. He argues that these layers change at different speeds and should not be tightly coupled.
The talk explores resumability, event and scheduled triggers, observability and sandbox state. A background-agent demonstration shows how an ongoing workflow can triage incoming work rather than waiting for a fresh chat session.
Dan Farrelly closes the architectural argument with feedback tied to actual user outcomes, not merely positive reactions to an answer. His claims about architecture half-lives are a design perspective rather than measured guarantees. Sponsorship, product invitations and booth promotions are omitted.
Watch on YouTube
