A new Internet-Draft filed with the IETF in March 2026, authored by engineers from Montcao and NVIDIA, proposes the Agent Identity Protocol (AIP) — a standard aimed squarely at a problem most enterprises still haven't solved: AI agents that inherit the full API access of the humans who deployed them, with no cryptographic line between what the human can do and what the agent is doing right now. The draft splits the problem into two layers. Layer 1 handles identity: every agent gets a unique identifier and a key pair registered with an AIP Registry, and it signs every outbound action with that key. Layer 2 handles enforcement — a proxy sits between the AI client and every tool server, verifies the signature, checks the request against a declarative policy, and returns an allow, deny, or hold before the tool is ever reached. The mechanics are specific enough to implement. Each outbound tool call carries an AIP Token, a compact signed JSON object that functions as the agent's assertion of identity to the proxy. The token binds to the exact arguments being passed — it includes a hex-encoded SHA-256 hash of the canonical JSON serialization of the tool call arguments, so a signed token can't be replayed against a different action. It also carries a 128-bit CSPRNG-generated nonce, unique per token, to block replay attacks outright. What makes this worth flagging isn't novelty of concept — signed, scoped, short-lived tokens are the same idea underlying most agent-identity proposals surfacing this year. It's the placement of enforcement. Rather than trusting an agent's own claims about what it's authorized to do, AIP puts a mandatory proxy in the request path itself, so policy evaluation happens as a gate, not an audit trail checked after the fact. That's a meaningfully different posture from frameworks that focus on post-hoc attestation: instead of proving later what an agent did, it tries to stop unauthorized calls from completing at all. The draft is still a draft — individual submission, no IETF endorsement, no working-group backing yet, and it expires in mid-September 2026 unless revised. Internet-Drafts at this stage are proposals, not standards, and plenty never advance. But the fact that NVIDIA is a listed co-author gives it more weight than the average solo IETF submission, and the timing tracks a broader pattern this year of infrastructure vendors trying to define the identity layer before someone else's spec becomes the default.