The guest argues that useful AI agents need a faithful map of an organization's processes, knowledge and dependencies. He suggests documenting those workflows before handing them to models and describes defensive automation that continuously checks the software and personal information he has exposed online. His warnings about future attacks and provider dependence are predictions and opinions, not demonstrated outcomes.
In a walkthrough of his personal agent setup, he shows how skills encode preferred outcomesAn AI agent skill is a reusable package of instructions, resources, and tool guidance for performing a bounded kind of work. rather than every instruction, while bookmarks, conversations and other inputs feed a shared context storeContext engineering designs how relevant instructions, evidence, state and tool information are assembled for a model's task.. An example turns captured material into candidate work items; he cautions that an agent should not automatically act on every idea found in a conversation.
The coding demonstration centers on one ideal-state documentSpecification-driven development defines intended behavior and acceptance criteria before implementation begins. that records a project's goals, decisions and granular criteria. The guest says those same criteria guide both implementation and evaluation, with deterministic assertionsDeterministic software verification checks an output with a repeatable procedure that produces the same result for the same inputs and state. and model judgments used where appropriate. He also shows task routing by required capability and costAI model routing sends each request to a model chosen for that request's complexity, cost, speed, privacy, or other requirements.. The durable practice is to keep the specification, generated code and checks in agreement while retaining human understanding of the system.
Watch on YouTube
