ZeroGate Proposes Trust-Preserving Fast Paths for Governed AI Agents

Summary

ZeroGate is a runtime design for moving some authorization work earlier in an AI agent’s dispatch path without treating that preparation as final authorization. An issuer signs a short-lived ActionPass for an exact action, while a trusted runtime adapter reconstructs the final action and a local gate verifies its binding and consumes its nonce. A SQLite transaction couples nonce consumption, applicable quota updates, and an admission receipt. The paper states a conditional decision-preservation proposition: local admission implies that a specified synchronous policy would authorize the same action at that point, assuming sound approval, complete and current policy dependencies, faithful observations, and atomic consumption. The implementation does not by itself guarantee fresh knowledge of the external world or exactly-once remote effects. The evaluation combines semantic fixtures, concurrency and crash experiments, and an Azure Blob study using the same issuer and gate. Across 4,800 cloud attempts, prepared worker-admission-to-dispatch p95 latency was 9.802–11.374 milliseconds, compared with 25.018–334.000 milliseconds for synchronous execution across the tested concurrency levels. However, prepared execution had a longer mean complete lifecycle at every level because preparation and batch dwell were included. The authors frame ZeroGate as an explicit revalidation contract and auditable authorization boundary, rather than a new cryptographic primitive or a universal performance improvement.