Representatives from PayPal and Google said at Consensus Miami that AI agents are structurally locked out of traditional bank accounts, leaving crypto rails as the natural payments layer. Google has launched the Agentic Payments Protocol (AP2), with 120 partners including PayPal, and donated it to the FIDO Foundation. A PayPal executive acknowledged that the open question keeping her up at night is "how do you onboard agents into all of the existing capital markets and infrastructure plumbing that powers payments and trading today."

AI agent blockchain transactions settled between May 2025 and April 2026, totaling over $73M, averaging roughly $0.31 per transaction (Keyrock, May 2026).

In one year, machine-to-machine payments went from concept to functioning on-chain ecosystem. The volume is real, the infrastructure is live, and major players (Coinbase, Stripe, Google, Visa), are actively building for this market. But a closer look at the numbers reveals an important caveat: the average transaction size of $0.31 sits well below the threshold where traditional payment rails become economically unviable, which means the current activity is largely experimental and crypto-native. Institutional participation remains limited.

The ecosystem is growing fast, but the gap between infrastructure buildout and real-world commercial demand is still very much open.

The past few weeks of research point to a question that sits underneath most discussions about agentic infrastructure: what does it actually take for a payment to be fast enough, private enough, and autonomous enough to serve machine actors?

Three articles published this month approach that question from different angles: latency, channel architecture, and peer-to-peer design. Together, they sketch a coherent picture.

Latency is not just a performance concern. When agents process retrieved data in fractions of a second, a payment system that takes two seconds to confirm becomes a structural bottleneck. The target the industry is converging on, around 400 milliseconds, is not arbitrary. It reflects the pace at which autonomous systems actually operate.

Payment channels remain the most efficient primitive for ultra-low-latency interactions. But their usefulness has historically been constrained by the cost and friction of channel management. A settlement layer that reduces that overhead by orders of magnitude changes the calculus: channels become viable for nanopayments, for pay-per-timestamp models, for interactions where the economics only work if overhead is near zero.

And underneath both of these sits a more foundational question: who gets to see what? Peer-to-peer payments, in the original sense, mean transactions that are validated by cryptography, not interpreted by intermediaries. In a world of continuous, automated, low-value machine interactions, that is not an ideological preference. It is a practical requirement.

Speed, efficiency, and neutrality are converging into a single design problem. The infrastructure that solves it will not look like an optimized version of what exists today.

A space to slow down, read deeper, and sit with ideas worth more than a scroll.

Stablecoins are becoming the preferred medium of exchange for agentic commerce, but most payment infrastructure still forces a tradeoff between privacy and compliance.

This piece explores how a payment layer built around client-side validation and zero-knowledge proofs can support both: full confidentiality by default, with the cryptographic tools issuers need to meet regulatory requirements.