$ today_in_10_seconds
Mastercard will now tell banks how likely it is that a payment was started by an AI agent. That's an inference, not a declaration, because agents still don't reliably identify themselves. Meanwhile OneTrust and Classie shipped runtime layers that sit between agents and the systems they touch. The checkpoint is moving out of the agent and into the infrastructure.
$ lead_story
Mastercard adds an "is this an agent?" score to Agent Pay
Mastercard has expanded Agent Pay with new trust services. The first is a probability score, in U.S. testing, estimating how likely it is that an AI agent initiated a transaction, so issuers can approve legitimate agent-led purchases. Mastercard plans to add behavior, merchant risk, credential risk and transaction-pattern signals over time. Cloudflare is contributing web-side signals in privacy-preserving environments, and Skyfire is supplying Know Your Agent verification. Mastercard's own example shows the problem. An agent books a flight, hotel and transport across several merchants with a single approval, which looks unusual on the surface. That is exactly the kind of pattern fraud models flag today. A probability score is what you build when agents won't say what they are. For builders, the takeaway is that how your agent identifies itself to the network now affects its approval rate.
// why_it_matters: An agent that can't prove what it is will get declined like a fraudster.
$ also_worth_attention
OneTrust CORIE puts policy enforcement at the tool call
CORIE enforces rules at the tool call by allowing an action, stopping it, or sending it for human review. The checking system stays separate from the agent, and it records the agent, request, purpose, outcome and governing policy as the work happens. The architecture is right. The catch is that most capabilities are planned for private preview, so for now this is a roadmap, not a product. OneTrust's own data shows the gap it's selling into: 87% of organizations encourage agent use, and only 47% have clear governance in place.
link →Classie Supervise ties agent control to agent spend
Classie Supervise is now generally available. It monitors agents and enforces policy across browsers, endpoints and enterprise compute, and it keeps a chain-of-custody record. It can alert, restrict, reroute or stop an agent action as it happens, and it attributes token spend to the users, agents and sessions behind it. The useful idea is putting cost and control in the same layer. A runaway agent shows up first as a security problem and then as a billing one, and here both appear in one view.
link →FP: the bottleneck is execution, not insight
FP Digital's report argues that the problem is a disconnect between insight and execution. It describes agents that act across ERP and CRM within boundaries set by human overseers, and it concludes that governance, not model capability, will decide who succeeds. The consultancy view now matches the vendor view, and the two stories above are selling exactly that governance layer.
link →
$ bigger_picture
The trend behind the headlines
Three vendors this week placed the same architectural bet in three different places: Mastercard at the payment network, OneTrust at the tool call, and Classie at the endpoint. Add this morning's Nvidia DPU enforcement (#107) and the VALUE broker (#108), and the pattern is clear: nobody trusts the agent to police itself. The weak spot is identity. CORIE and Classie can govern agents they can see. Mastercard has to estimate the odds that an agent is even involved. Until agents carry declared, verifiable identity from end to end, every enforcement layer is partly guessing.
Skyfire
Skyfire provides Know Your Agent technology for AI agent identity, verification and payments, and it now plugs into Mastercard Agent Pay. If your agents transact, it's a practical way to see what presenting verified identity to issuers and merchants looks like before the network has to guess.
$ subscribe --free
Get the next edition in your inbox
// DAIly Agent covers agentic AI three times a week
// free, no spam, exit anytime