Midnight says a USDM-branded ATM is live in Midnight City for agents to test on its Preview Network. The careful reading is test infrastructure in a preview environment, not a production payment system. The supplied excerpt establishes the machine and intended agent testing; it does not establish transaction volume, autonomous customer use or financial performance.
A separate Midnight statement says Policy Vaults could give AI agents predefined risk limits enforced in real time, with zero-knowledge proofs making compliance verifiable without exposing proprietary logic. The word could matters. The post describes a proposed design direction, and the briefing itself says the idea may still be at a concept or demonstration stage. It should not be reported as deployed control.
Other items in the lane are useful mostly as boundaries. Polygon's cited message concerns continuous onchain availability rather than a ZK release, while the briefing says its recent attention also included flood-relief fundraising. Ronin material is described as game-event coverage. Their presence in the reviewed material does not provide evidence about zero-knowledge engineering and should not be used to inflate the category.
The core operating problem is how an agent proves it stayed inside a rule without exposing the entire strategy. A proof can help attest to a defined condition, but the policy still needs an owner, sound inputs and a response when execution diverges. Privacy cannot substitute for accountability, and autonomous decision-making is not itself a safety model.
Xai's documented gas-subsidy path uses ERC-2771 meta-transactions and EIP-712 signatures so a relayer can sponsor player gas. In a hypothetical agent system, explicit signed authorization and sponsored execution could be evaluated alongside private policy proofs. This comparison does not indicate that Midnight uses Xai, or that Night Ash operates an agent, payment rail or deployed Xai integration.
The next evidence should be a constrained trial with published rules, failure cases and independently inspectable results. If a system can reject actions outside policy while preserving necessary confidentiality, the concept can advance toward credible control. If proofs cover only a narrow condition or the input source remains manipulable, operators should revise the design before assigning real authority.