Every agent protocol built so far answers the same question: how do agents talk to each other? Model Context Protocol standardises tools. A2A standardises discovery. They’re good at plumbing, transport, capability, task lifecycle. But notice what none of them ask.
Who authorised this? Under what authority, and for how long? Within what budget? Who has to approve it? And when it’s over, what can you actually prove?
That’s not a small gap. Researchers call it a missing architectural layer. Six hundred and seventeen governance findings across six popular agent stacks. Roughly two thousand MCP servers surveyed, uniformly unauthenticated. And the EU AI Act’s event recording obligations became enforceable this August.
Correction That line is wrong. The GRIP paper gives the dates when the EU AI Act’s record-keeping rules for high-risk systems (Article 12) apply: 2 December 2027 under Annex III, and 2 August 2028 under Annex I.
GRIP is the Governed Runtime Intent Protocol. A governance contract for agentic work. Here’s what it binds into one agreement.
An intent, the unit of work, written as a contract. What must be true when this is finished. What constraints hold. What counts as proof. And a budget with four axes: input tokens, output tokens, time, and money.
Authority, as short-lived signed leases. Not a password. Permission to do one thing, for one purpose, until it expires. The root signing key lives outside the engine, so the runtime can verify authority but can never mint it for itself.
Attention, a governed queue where every refusal is typed, and carries its reason in plain language. Context, compiled deterministically, addressed by digest, with a receipt that explains what was left out, not just what went in. Execution, with effect boundaries, so a retry cannot fire the same real-world action twice. Settlement, evidence, checked criterion by criterion, signed. The state machine itself refuses to close an intent without proof.
And waitpoints, human approval as proof of possession: the engine holds only the digest of your approval token. It is structurally incapable of approving on your behalf. Every gate decision lands in a digest-chained ledger. Which makes a GRIP runtime two things at once: auditable, for compliance, and trainable against.
MCP and A2A describe what an agent can do. GRIP contracts what it may do, what it owes, and what it can prove. The protocol is specified as typed messages, and implemented in a reference runtime released together with the paper. This is GRIP’s debut.
Correction The runtime was not released with the paper. Foxora Runtime is in private preview. What came out with the paper is the specification and its files: the schemas, the registries, a verifier and seven test vectors.
The figures the film gives about other systems come from research the paper cites.