Guillermo Rauch and David Ondrej argue that human judgment shifts toward goals, taste and the constraints around agent-written software. Rauch distinguishes iteration speed from useful progress and recommends personal or company evaluations instead of adopting prompting advice solely because it is popular.
Guillermo Rauch describes a CLI-centred workflow with research and project directories, reusable agent instructions, model experimentation and an embeddable harness. The conversation includes interface-generation examples and explores how cloud sandboxes, shared filesystem state and production feedback can support longer-running work.
Guillermo Rauch connects verification to measurable system behaviour, including latency, error-free sessions and observable traces. The discussion covers security review, memory-safe implementations and containing supply-chain risk; it presents architectural recommendations and company demonstrations, not proof that agents or cloud sandboxes eliminate vulnerabilities.
Watch on YouTube
