Alexander Embiricos and Romain Huet frame engineering as problem solving rather than only writing code. They trace a progression from completions and inline changes to agents that test work and pursue longer goals, while emphasizing useful product decisions and collaboration.
Their product-design discussion combines conversational direction with a hands-on interface for inspection and steering. Huet describes the model API, open harness, app server and plugin layers as shared infrastructure used by OpenAI's own products and available to developers. This is the presenters' architectural account, not a claim that every product feature is interchangeable.
The talk discusses cost efficiency, faster inference and the desire to move isolated tasks beyond a permanently open laptop. Model comparisons and announced speeds are attributed to the presenters, and future local-cloud integration is distinguished from functionality already demonstrated.
Peter Steinberger describes moving from polling many terminals to directing a long-running manager that delegates work. Persistent context, coordination and triggers support his proposed loop; workers investigate, implement, test and review before returning evidence for human decisions.
Steinberger argues that attention becomes the next constraint after tokens and compute. The inner execution loop can be delegated, while people set direction and review meaningful outputs. He explicitly says the seamless cross-machine manager is not yet solved, leaving harness and organizational design as continuing engineering work.
Watch on YouTube
