Agent governance paper shifts oversight from actors to actions

A new cs.AI paper, “Governing Actions, Not Agents,” argues that autonomous AI systems should be governed through evidence attached to consequential actions, not by trying to continuously inspect their internal reasoning.

The authors formalize institutional attestation as a computational governance model for agents that may eventually perform irreversible tasks such as clinical prescribing or production software deployment.

The idea is timely because agent systems are moving toward real-world actions, where auditability and external proof may matter more than a model’s stated chain of thought.