A new paper proposes a fix for a gap most agent protocols don't address: when one AI agent calls another on a human's behalf, it accepts the callee's terms of service, but nothing in the protocol proves it understood those terms or actually followed them afterward. The author, independent researcher Ravi Kiran Kadaboina, names this the accountability gap. Existing auth layers like OAuth and mutual TLS establish who is allowed to call which capability, but they don't address under what conditions a permitted call may be made, and those conditions shift as the callee's policy changes over time. Anumati (Sanskrit for consent) draws a line between two kinds of proof. "Proof of acceptance" is what we already have — a timestamped click-through acknowledgement, borrowed straight from human web forms. "Proof of adherence" is new: a per-action reasoning record that cites the specific clause an agent evaluated before acting. Kadaboina argues the inherited human-authentication model — a timestamped acceptance record — is inadequate for autonomous agents because it proves acceptance but cannot prove adherence. The mechanism is three primitives. A PolicyDocument holds the callee's versioned terms. A ConsentRecord captures the calling agent's acknowledgement of a specific version. An AdherenceEvent logs each individual action against the clause it was checked against, chained together into an append-only audit trail. The pitch for why this matters now: agents can do something humans structurally can't — parse a policy document, evaluate its clauses at runtime, and generate a reasoning trail for every single action taken under it. Anumati's primitives exist to make that capability legible rather than invisible. Critically, the model is designed as a non-breaking extension to two protocols already in wide use — Agent2Agent (A2A) and Model Context Protocol (MCP) — rather than a rip-and-replace standard, with a TLA+ specification and a reference Python implementation released alongside the paper for validating chain integrity. The open questions are real. The scheme assumes calling agents can reliably generate detailed, clause-citing adherence records in real time, and that callee policies stay stable enough to reference without breaking the chain. The paper doesn't include performance overhead measurements or tests against agents whose policies shift mid-session, so how this holds up under production load and volatile terms of service remains untested. Still, it's a concrete attempt to formalize something the field has mostly waved at: not just who an agent is, but whether it actually did what it agreed to.