Blockchain Agent to Agent Payments
This is the short version. The essay it summarises has been years in the writing and is still being edited. When it is finished it will stand here in full.
A landmark paper by eleven researchers has identified one of the most important near-term problems in AI infrastructure, and its diagnostic work is the best public articulation of that problem I have read. Agents acting continuously across hundreds of interactions cannot pause for human authorisation at each step, and payment infrastructure built for humans does not extend to that world.
Its central contribution is a four-stage lifecycle: discovery, authorisation, execution, accounting. The failure taxonomy at each stage is sharper still, and the most important finding is named precisely. Payment completion is only weakly coupled to service completion, so a provider can take valid payment and deliver nothing while the settlement layer records the payment as final.
The architecture the paper recommends cannot solve the problem it has just documented. Sequencers are specific companies in specific jurisdictions that can include, exclude, delay and reorder transactions, with no mechanism by which a user can compel one to process anything. Account abstraction adds three further intermediaries to every payment flow. The paper spends fifteen pages documenting an accountability problem and then proposes infrastructure that multiplies the number of parties who must be trusted.
The mechanism that resolves the coupling failure has eleven years of peer-reviewed literature behind it and appears nowhere in the paper's ninety references. A systematisation that maps the architectures with corporate sponsorship and overlooks those without it is a map drawn by the cartographers the territory paid.
