A new research system called EvidenceNet checks whether AI agents managing network configurations actually finished the job, not just whether their commands ran without error.
Operational networks span many devices and administrative domains, so no single AI agent has full visibility or authority. One agent might issue a configuration change while others, restricted to their own scopes, cannot see whether it worked network wide. The paper calls this the completion admission problem: a command executing successfully does not prove the network reached the intended state, and that state can go stale the moment something else changes. EvidenceNet adds a broker that collects post change observations from the required scopes, an admission gate that checks evidence is current and properly sourced, and a verifier agent that reviews the content. Tests on live routing networks showed it caught successful outcomes that configuration logs alone missed, and correctly rejected completion when evidence was stale, swapped, or pulled from the wrong source.
The interesting move here is procedural, not algorithmic. As companies push more network operations onto AI agents, "the command executed" quietly becomes the working definition of success, even though nothing confirms the network itself changed. EvidenceNet treats an agent's report as a claim to verify rather than a fact, the same instinct that pushed distributed systems toward consensus protocols and audit logs instead of trusting one node's word.
It is less a new capability than a seatbelt for one already being installed industry wide.