Attestation
Beyond the Click-Through: A Proposal for Agents That Prove They Followed the Rules
Most agent-to-agent authentication answers one question: who is allowed to call what. It says nothing about whether the calling agent actually understood or honored the callee's terms once the call went through. A new paper, Anumati, from independent researcher Ravi Kiran Kadaboina, names that gap directly.
The paper distinguishes between two kinds of proof that get conflated today. Proof of acceptance is a timestamped acknowledgment — the agent equivalent of clicking "I agree." Proof of adherence is something stronger: a per-action record citing the specific policy clause an agent evaluated before acting, the decision it reached, and the reasoning behind it. As the paper puts it, authentication protocols like OAuth and mutual TLS establish who may call which capability, but they don't address under what conditions a permitted call may be made, and those conditions change as the callee's policies evolve.
The design rests on three primitives. A PolicyDocument is the callee's machine-readable usage policy, versioned and content-addressed. A ConsentRecord captures the calling agent's parsed understanding of that policy, clause by clause. An AdherenceEvent is logged for every action, tying it back to the specific clause the agent evaluated. Chained together, these form two append-only lists — a consent chain and an adherence trail — that satisfy five properties the paper lists as relevant to legal and regulatory accountability: completeness, traceability, tamper evidence, version fidelity, and optional ledger anchoring.
What makes this more than a theoretical framework is that it's built to slot into protocols already in production use. The author instantiates the model as a non-breaking extension to both the Agent2Agent (A2A) protocol and the Model Context Protocol (MCP), using each protocol's existing extension mechanism rather than requiring a fork or a new standard. The paper also includes a TLA+ specification of the consent lifecycle and a reference Python implementation for validating chain integrity and adherence trails, which at least gives outside reviewers something concrete to test rather than just a diagram.
It's one researcher's proposal, not an adopted standard, and independent reviewers have flagged that performance overhead at high call volumes is still unmeasured. But the framing is useful regardless of whether Anumati itself gets adopted: it separates "the agent said yes" from "the agent actually behaved as agreed," a distinction that current agent protocols mostly skip over.