Viren Baraiya argues that a production agent harness is an application, not just an LLM loop. It may coordinate background work, events, humans and multiple specialized agents over long periods, so infrastructure failures must not erase progress or completed side effects.
The architecture separates reasoning from execution. The model proposes what to do next, while deterministic code decides how an action is performed, applies required human gates and records whether it has already happened. Viren Baraiya connects this pattern with durable workflows and late-bound sagas.
A prepared SRE demonstration compiles an agent's proposed steps into a Conductor workflow, investigates a problem, performs a rollback and verifies recovery. A later iteration recognizes the rollback already happened. The example illustrates planning combined with predictable execution; it is a staged demonstration rather than independent evidence of production reliability.
Watch on YouTube
