agentic-paybench
Start here. Everyday AI Ltd, a UK research company, measures the payment rails that software agents use to pay each other. Each method, frozen and anchored before any of its results was published, is applied identically to every rail; no rail is recommended and no ordering is published. This index links what exists: the method and its harness, the signed capability records for six rails, the two published measurements, and the three public pre-registrations.
Research artefacts for measuring agent-to-agent payment rails: a pre-registered evaluation methodology and a capability oracle. Rail-neutral by design: the method is designed for pairwise comparison, neither published measurement asserts a merit ordering of named rails, and the work holds no commercial relationship with any rail it measures.
| Repository | What it is |
|---|---|
| payhelm | PayBench: the measurement methodology (frozen v1.2), harness, and calibrated fixtures, as a fork of HELM. Pre-registered and anchored before publication of any result: OSF DOIs 10.17605/OSF.IO/XGFUJ, 10.17605/OSF.IO/UFQG5 and 10.17605/OSF.IO/UETFW (the test-network confirmatory data pre-registration, registered 2026-10-15, on top of the v1.2 methodology). |
| oracle | The capability oracle: schema, ingestion, methodology, and the static read surface, live at oracle.agentic-paybench.dev. |
What is published, and what is not. The methods are public. The authorization-primitive latency (DIM-02) findings for six named rails and the settlement-finality (DIM-01) measurement of five named rails on test networks are published as measurements at oracle.agentic-paybench.dev/results/. No league table is published: the Bradley-Terry strengths and the pairwise win matrix the settlement-finality method computes are withheld by choice, not omission. No rail is recommended here, and no payment is routed.
Posture
These figures are a measurement, not a recommendation: no rail is endorsed, no ordering implies a best buy, and no rail is being suggested for use.
Rails are listed alphabetically by rail name. That ordering is fixed and carries no performance meaning.
The same measurement protocol was applied identically to every rail.
Rails covered by the methodology: x402, AP2, the Machine Payments Protocol (the Stripe and Tempo HTTP 402 protocol, on more than one settlement rail), and Lightning-based rails, measured on test networks with the firm's own traffic.
This page is the research index for the agentic-paybench GitHub organisation. Nothing here is a product, a service, or an invitation to transact.