New AI governance model uses institutional attestation to control autonomous agents

A new paper titled "Governing Actions, Not Agents" from Jakob Salfeld-Nebgen introduces a computational governance model that takes inspiration from how human institutions oversee powerful actors—by focusing on the actions themselves, not internal reasoning. As autonomous AI agents begin to perform irreversible, consequential tasks like prescribing medication or deploying production code, existing safety approaches that attempt to monitor or constrain the agent's reasoning become impractical and brittle. Instead, the proposed model separates planning autonomy from execution authority.

Under this system, an AI agent can freely plan and reason about any course of action. However, for designated high-risk actions, execution is conditional on a set of preconditions that must each be independently attested by a separate authoritative source. These attestations are cryptographically bound to a declared intent and evaluated by a deterministic policy. Every decision is recorded in a tamper-evident log that anyone can independently re-verify. The paper includes a proof-of-concept implementation and illustrates the model with detailed examples from software deployment (requiring code review attestation before pushing to production) and clinical prescribing (requiring licensed physician attestation before dispensing a drug).