Abstract

Abstract (Part 7 of 27)

AI Agent Safety: The Evaluator Picks Who Acts Next — But Cannot Let Them Act

A valid negotiated plan and a favorable risk score still do not identify which of several independently controlled agents should safely take the next step. Concurrent agents may contend for the same resource. An agent chosen by a supervisory model may lack the right local credentials. A previously permitted agent may become unavailable, disqualified, or unsafe after new feedback. Two organizations can disagree about whether the selected actor is eligible. An evaluator’s natural-language direction can be treated as unconditional authority, and a decision not tightly bound to the session, frozen plan, phase, epoch, and precise next action can be replayed or redirected.

This embodiment (Part 7 of a 27-part series; full details at https://zenodo.org/records/23940766) specifies the actor-selection and instruction-routing layer. After the joint plan is committed, frozen, and evaluated, the independent evaluator screens candidate actors against local eligibility rules, hard bans, risk evidence, and resource conflicts. It may select one agent, a subgroup, a qualified substitute, a human reviewer, or no one. The resulting directive is typed, context-bound, and non-effective: it identifies the permitted next turn and restrictions but does not constitute an execution credential. A separate organizational next-step issuer confirms that current local policy still permits the precise actor and instruction. Each domain’s protected gateway or Finality Sink independently validates the live pending message or operation, recipient, resource, expiry, revocation status, and single-use consumption before admitting anything. During this process, governed channels remain restricted. Recovery paths handle unavailable actors, contradictory policies, partitions, and missing receipts without silently restoring rights.

The architecture therefore routes the next authorized opportunity in multi-agent systems while ensuring that evaluator selection alone never becomes permission to communicate or act.

Keywords: multi-agent systems, agentic AI, AI safety, AI governance, next-actor selection, evaluator-directed continuation, controlled next-turn instructions, non-effective directive, fail-closed systems, zero-trust agents, cross-organizational collaboration, actor eligibility, local veto, protected scheduling, execution finality, identity-and-operation-bound continuation, separately enforced continuation, candidate actor screening, non-bearer next-step authorization, replay-resistant actor selection, multi-agent turn control, independent next-actor evaluation, protected actor routing, eligibility disagreement resolution, non-effective next-move assignment.

Creative Commons License

Creative Commons License
This work is licensed under a Creative Commons Attribution-Noncommercial 4.0 License

Download

Share

COinS