Rails for the Swarm: Why Structured Workflows and Human Judgment Matter More in the Agentic Era
Mentat argues that LLMs and agentic workflows make structured DevOps tooling and human judgment more important, not optional. Agents excel at drafting code and summarizing failures but lack organizational memory and scar tissue from past outages, which creates unowned systems when generation outruns governance. Mentat's recommended pattern is deterministic rails around non-deterministic reasoning—Infrastructure as Code, policy as code, observability, provenance, and real human review—so agent speed becomes transportation rather than a convincing accident. Continuity with Mentat's foundations-before-the-penthouse thesis: agents accelerate whatever rails you give them, and humans remain the ones who can feel when the destination is wrong.