How to Know When an AI Agent Can Work Unsupervised — Evan Kormos, Coinbase
Coinbase engineer frames organizational policy as the original system prompt for governing AI agents
“Think of policy as the original institutional intent or the original system prompt.”
Evan Kormos of Coinbase argues that existing organizational policy documents—written with terms like 'must,' 'may,' and 'should'—serve as the foundation for governing AI agents in enterprise workflows, effectively acting as a system prompt. He proposes separating build-state policy from run-state agent behavior and using directed acyclic graphs to enforce gated, parallel workflows while maintaining compliance checks. The insight is practically useful for teams deploying agents in regulated industries but represents practitioner wisdom rather than a novel research or product announcement.