Most agent-to-agent authentication answers one question: is this caller allowed to invoke this capability? A new paper called Anumati argues that's the wrong question to stop at. When an agent calls another agent's service, it typically accepts the callee's terms of service the way a person clicks 'I agree' — a rubber stamp, not a demonstration of understanding or compliance. The paper's author, Ravi Kiran Kadaboina, draws a formal line between two things that get conflated today. Proof of acceptance is just a timestamped acknowledgement — the agent equivalent of clicking through a EULA. Proof of adherence is different: a per-action record showing which specific clause of the callee's policy applied to that action and how the calling agent reasoned about it. Authentication protocols like OAuth and mutual TLS establish who can call what, but they don't address the conditions under which a permitted call should actually happen, especially as the callee's own policies change over time. To make this concrete, the paper defines three building blocks: a PolicyDocument holding versioned terms, a ConsentRecord capturing that the terms were accepted, and an AdherenceEvent that logs the reasoning behind each individual action against those terms. Together they form an append-only, versioned trail rather than a one-time checkbox. The design is proposed as a non-breaking extension to two protocols already seeing real adoption — Agent2Agent (A2A) and Model Context Protocol (MCP) — meaning it's meant to bolt onto existing infrastructure rather than replace it. The author backs the design with a TLA+ formal specification of the consent lifecycle plus a reference Python implementation for validating chain integrity and adherence trails. What's notable is the gap this targets. As agents increasingly delegate to other agents on a human's behalf, the terms a caller agreed to at time zero may no longer match the callee's current policy, and there's currently no protocol-level way to show a specific call actually honored whatever policy was in force at that moment. Anumati's per-action, clause-citing record is an attempt to make that traceable rather than assumed. The proposal is early — a single-author arXiv paper, not a ratified standard — and independent reviewers note it doesn't yet measure the overhead of generating detailed adherence records in real time, or test how the model holds up when a callee's policy shifts mid-session. Those are real open questions before any implementer adopts it. But it names a genuine hole in current agent protocol design: consent that expires the moment it's granted, with nothing tracking whether it was actually kept.