AI Agents Are Not Traditional Software
Why agents act through ongoing interpretation in real time, and the case for a provider-agnostic authority layer.
Traditional software executes instructions. An agent interprets intent, in real time, against a world that keeps moving. That difference is not a matter of degree, it changes what it means for the system to be correct.
A deterministic program can be verified once and trusted thereafter. An agent's behavior is produced fresh at each step from a model, a context window, a set of tools and an accumulated history. Any of those can change without the credentials it holds changing at all.
This essay argues that consequential agents require an authority layer that is independent of any single model provider: authority that originates with a person or organization, is bounded, is revocable, is checked at the moment of action, and leaves evidence behind.
It closes with the Authority Checklist for Consequential Agents, a practical set of questions for teams putting agents in front of money, customers or infrastructure.