A new arXiv paper called Anumati — Sanskrit for consent — targets a gap nobody's protocol has closed: when one AI agent calls another on a human's behalf, it clicks through the callee's terms of service the way a person clicks through a EULA, but that click proves nothing about whether the agent actually behaved within those terms afterward. The paper's author frames this as a structural accountability gap. As the abstract puts it, the calling agent accepts the terms of service of the callee "without any protocol-level mechanism to prove that it understood those terms or that it subsequently honoured them." Existing authentication layers don't help here — OAuth and mutual TLS establish who is allowed to call which capability, but they say nothing about the conditions under which a permitted call may actually be made, and those conditions shift as the callee's own policies change. Anumati's core move is a formal split between two kinds of proof. One is proof of acceptance: a timestamped acknowledgment, the digital equivalent of a checkbox. The other is proof of adherence: a per-action reasoning record that cites the specific clause the agent evaluated before acting. The paper argues the first kind — inherited straight from how humans authenticate to services — is structurally inadequate for agents, because it proves acceptance but not follow-through. To make adherence provable, the paper proposes three primitives: PolicyDocument, ConsentRecord, and AdherenceEvent, which together form a versioned, append-only consent model for agent-to-agent traffic. The reasoning here leans on something agents can do that humans generally can't at scale: parse a policy document, evaluate its clauses at runtime for every single action, and emit a reasoning trail each time. A linked-list audit chain built from those events is meant to give the calling agent's human principal actual legal standing if something goes wrong downstream. Notably, the model isn't proposed as a new standalone protocol requiring adoption from scratch — it's designed as a non-breaking extension layered onto two protocols already in wide use: Agent2Agent (A2A) and the Model Context Protocol (MCP). The author backs this with a TLA+ specification of the consent lifecycle and a reference Python implementation for validating chain integrity and adherence trails, both published alongside the paper. The distinction matters because most of the industry's current consent tooling — cookie banners, API terms, click-through EULAs — was built for humans who accept once and then act unpredictably. Agents accept once and then act deterministically, at machine speed, thousands of times a day. Anumati is a bet that the record of those individual actions, not the initial click, is where accountability actually has to live.