Attestation
A Proposal for Post-Deployment 'Cards' That Track What Agents Actually Did
Most agent-governance proposals focus on identity at launch: who deployed this agent, under what credential. A new article from AI Frontiers argues that's only half the problem, and proposes a second instrument aimed at the other half — what happens after deployment.
The piece, by Gillian Hadfield, lays out two complementary infrastructure pieces. First, agent ID and registration systems, which let harms be traced back to responsible parties. Second, and less discussed elsewhere, what the article calls 'model deployment cards' — reports that track agent behavior in the field after release, rather than static documentation produced before it ships.
The distinction matters because most existing 'card' culture in AI — model cards, system cards — describes a model's training data, intended use, and known limitations at release time. They're snapshots. An agent that's been operating for six months, chaining tool calls, negotiating with other agents, and acting on live infrastructure has generated a behavioral history that a pre-deployment card can't capture. The proposal is to create a living record of what the agent has actually done, intended to create visibility into real-world impacts that current documentation standards don't provide.
The framing draws urgency from a concrete failure mode: incidents where attribution breaks down after the fact. The article points to the case where agents behind the Hugging Face breach were eventually traced back to OpenAI, but notes that future attacks could be harder to attribute to an accountable party — for instance if a malicious actor deploys agents while covering their tracks, or if an AI without a clear human principal runs itself on poorly monitored infrastructure. In those scenarios, parties harmed by an agent have no one to pursue, which the article calls a governance gap rather than just a technical one.
What's notable here isn't the ID half of the proposal — registries and agent identifiers are now a crowded field, with government, standards-body, and vendor proposals all competing. It's the second half: a push to standardize *post*-deployment reporting as its own artifact, separate from both the registration record and the incident log. Whether 'deployment cards' becomes a term of art the way 'model cards' did, or gets absorbed into existing audit-trail and observability tooling, isn't settled. But the proposal names a real gap — the absence of any standard format for documenting an agent's accumulated behavior over its operational life, as opposed to its specs at birth.