Zero Framework Cognition

A late-2025 architectural note that started life as Steve shouting at every team he was building with: "no! bad! when we reach a decision point, send it back to the model. Don't try to hack this on the client side with regular expression matching!" The post is the codified form of that shout. Four AI-app builds in, the anti-pattern is consistent enough to name: engineers reaching for client-side keyword matching, regex, or hand-rolled state machines to decide the agent's next move. The post calls the alternative Zero Framework Cognition: the client owns no cognitive load; every branch point routes back to the model. The cost is a few extra tokens; the payoff is an application that doesn't quietly drift back toward the same fragile if-then-else code it was trying to retire.

A craft note for anyone building on top of LLMs. Read with Six New Tips and The Future of Coding Agents as a small late-2025 set on what holds up after a year of agentic building.