Agentic Payments Field Manual

Discovery got solved by MCP and A2A. Payment got solved by x402, USDC and smart wallets. Identity is being solved by Skyfire and the card networks. Work-quality verification is the one layer with no production answer — and it's the layer that touches every single transaction, which is what makes it the defensible one.

The structural reason it stays empty is worth remembering: payment rails earn on throughput, and a quality gate reduces throughput. Circle, Stripe and Coinbase all make money on volume, so none of them wants to be the party that says "this doesn't settle." That misalignment — not the difficulty of the engineering — is what keeps the slot open.

AI agents are the first economic actor that cannot use the existing payment system. Every rail we have — cards, ACH, invoicing, API keys — assumes a human signs up, holds an account, agrees to terms and disputes charges. An agent doing autonomous work needs to pay a service it has never met, for an amount too small to justify onboarding, at machine speed, with no human in the loop. That's a structural mismatch, not a marketing gap.

In March the story was "one standard is winning." By September the stack had decomposed into distinct layers with different winners in each. That fragmentation is better for a neutral trust layer, not worse — but it killed the "just plug into x402" simplicity.

The competitive read that held up. Skyfire, Crossmint, ChaosChain, Ritual and Coinbase are not lanes of one race — they're different geometries. Skyfire proves who the agent is. Crossmint aggregates protocols one layer up. ChaosChain proves settlement executed correctly. Ritual and EZKL prove the claimed model actually ran. None of them answers was the work good.

The most surprising finding of the research was how much of the stack is genuinely production-ready — not testnet, not "coming soon," but usable with real money. And a matching list of things that sound imminent and aren't.

This is the part of the research worth keeping longest, because it generalizes well beyond crypto. Any system where one party pays another for judgment-dependent work runs into the same three questions — and they are constantly conflated.

Three axes people collapse into one

Separating these is what made the competitive landscape legible. "We do verification" is meaningless until you name the cell.

The tiered cascade — and why it's economically viable

No single tier works. The cascade escalates from cheap-and-narrow to expensive-and-expressive, and only when confidence demands it. Blended across 10,000 tasks/day this lands near $0.68 per task — roughly 22× cheaper than all-human review.

Where LLM-as-judge breaks down

The headline "80–85% agreement" is a blended average, and the blend hides everything that matters. Agreement is a function of whether ground truth exists at all.

These were recorded as hypotheses, not measurements. Published category-level benchmarks were never found.

Six ways a judge gets fooled

The infrastructure was built ahead of the demand. That's the window the whole thesis rests on — and it's also the risk. These are the numbers that cut against the story, kept deliberately because they're the ones that get forgotten first.

Three things that broke the March assumptions

Solana flipped Base. The M1 report assumed Base as the launch chain. By mid-January 2026 Solana had overtaken Base on daily x402 volume and ran roughly 65% of year-to-date transactions. Base still led cumulative ($21.5M vs $16.4M) and enterprise tooling, but the cost gap was real. The lesson recorded at the time: build chain-abstract from day one.

The valuable category is the smallest one. Where x402 dollars actually flowed: agent-to-agent services $548.5k, infrastructure and utilities $267.1k, AI-generated services $14.2k. The category where quality verification matters most is the one with almost no volume — either because demand isn't there, or because verification cost is what's gating it. That ambiguity was never resolved.

The judge doesn't pay for itself at micropayment scale. At $0.01–$0.15 per evaluation against a $0.20 average transaction, verification eats 5–75% of transaction value. The same product is either a non-starter or trivial overhead depending purely on whether agent transactions grow past ~$5. That is the single load-bearing economic assumption in the entire thesis.

How to read an adoption announcement

The most portable habit from the later research. Four distinct stages, constantly collapsed into one in press coverage: an announcement → a usable product → an executed transaction → sustained usage. Visa naming Cross River and Lead Bank as settling in USDC is stage three for a specific corridor, not whole-bank adoption. Western Union's USDPT launched while several of its services were still being built. Most of the market sits at stage two. Apply the same ladder to legal authority: a bill introduced, a bill enacted, an agency rule, and a staff statement carry very different weight — the SEC's own 2025 liquid-staking staff statement says outright that it has no legal force.

Five questions were logged with explicit priors rather than answers. They're recorded here as they stood, because a stated prior is worth more later than a vague recollection.