Brazil Built the Best Non-Crypto Agent Rail and Put No Agent Protocol On It

Brazil has the payment rail an autonomous agent would design for itself: public, account-to-account, instant, always on, addressable by API, free for consumers, reported to carry more than half of all financial transactions in the country, and — since June 2025 — equipped with a native recurring mandate. Brazil does not have agentic payments. Both sentences are true on 3 August 2026, and the gap between them is the whole subject of this page.

The rail, in Banco Central numbers

Pix entered a limited test on 3 November 2020 and launched fully on 16 November 2020. The BCB-attributed annual series runs 9.4bn transactions and R$5.2tn in 2021, 24.0bn and R$10.9tn in 2022, 42.0bn and R$17.2tn in 2023, 63.4bn transactions and R$26.4 trillion in 2024. Through October 2025 the system had moved R$28tn, with R$3.32tn in October 2025 alone.

Two figures fix the scale better than the annual totals. On 5 December 2025 Pix settled 313.3 million transactions worth R$179.9 billion in a single day, beating the 297.4 million recorded a week earlier on Black Friday. At end-November 2025 the register held 178.9 million users — 162.3 million individuals and 16.6 million companies — in a country whose 2024 population the World Bank puts at 212.0 million.

One caution about the user counts, because it is the kind of thing this wiki exists to catch. Agência Brasil, in the same month, reports "170 million adults and more than 20 million companies" using Pix, and attributes that to the Banco Central too. The two BCB-attributed pairs do not reconcile: 170m against 162.3m individuals, and 20m+ against 16.6m companies. We do not know which definition each uses — accounts versus people, "uses Pix" versus registered in the key directory — and neither source says. We quote the itemised pair because it is itemised, and flag that a widely repeated "20 million companies" is 20% above the number the same institution is reported as publishing three weeks later.

Full-year 2025 is reported at roughly 79.8 billion transactions and around R$35 trillion, and Pix is reported to have carried 54.7% of all Brazilian financial transactions in the second half of 2025. We mark those as reported, not verified: we found them in vendor analysis rather than in a BCB release we could fetch, and the vendor in question sells into the market the number flatters.

A comparison with x402: HTTP 402 Finally Gets a Job, at 32 Cents a Transaction is tempting and we will make it, with the label attached: that page reports roughly 75 million x402 transactions in a trailing month averaging about 32 cents, and its own provenance describes that number as "a counter published by x402.org itself, with no stated methodology and no independent audit", re-quoted by CoinDesk from x402.org and therefore "NOT independent corroboration". So this is an order-of-magnitude contrast between a verified number and an unverified one: Pix's single record day is several times x402's entire reported month, and on the reported 2025 total Pix would average around 219 million transactions a day, every day, since before agentic payments were a phrase. The direction of the gap is not in doubt. The ratio is not something we would defend to a decimal.

Cards lost the checkout, and that is the part that is verifiable

Reuters reported on 10 February 2026 that Pix took 42% of Brazilian e-commerce in 2025 against 41% for credit cards — the crossover — and that consumer-to-business payments were 46% of all Pix transactions in January 2026 against 40% person-to-person. Pix stopped being a transfer app and became a checkout.

The forward number needs a discount. The projection of 45% by end-2026 and 50% by 2028, with the lead over cards widening to 14 percentage points, is Ebanx's, built on PCMI data. Ebanx is a cross-border processor whose revenue depends on Pix acceptance; this is a vendor forecasting growth in the product it sells. We tried to corroborate the 2025 share against Worldpay's Global Payments Report through a secondary write-up; that article had been retracted on 12 April 2026 "pending source verification review", so we dropped the corroboration rather than launder it. The 2025 crossover stands on the Reuters report alone. The 2028 half-the-market figure should be read as a seller's forecast that may well come true.

Pix Automático is a mandate, and nobody says how big it is

On 16 June 2025 Pix Automático became mandatory on the payer side, by regulation rather than by adoption. Resolução BCB nº 402 of 22 July 2024 inserts into the Pix rulebook an Art. 11-S reading "Todos os participantes provedores de conta transacional devem disponibilizar o Pix Automático para seus clientes, nos casos em que estejam atuando como prestadores de serviço ao usuário pagador", sets the date in Art. 2 — "a iniciação das transações de pagamento relacionadas ao produto Pix Automático e suas demais funcionalidades a partir de 16 de junho de 2025" — and attaches a daily fine, capped at sixty days, for any participant that has not been homologated and live by that date. The obligation is asymmetric, and the asymmetry is worth keeping: Art. 11-T states that "A oferta de Pix Automático para usuário recebedor é facultativa". Every account provider must let its customers *pay* this way; nobody is compelled to let its customers *collect* this way. Read its mechanics next to the ones in A Spending Limit Is Not a Conduct Limit: the payer authorises once; debits then run automatically; the payer sets the frequency, the amount (fixed or variable) and the maximum per transaction; the authorisation is cancellable and adjustable at any moment; it works only with individuals as payers and companies as collectors. That is, feature for feature, a spending mandate — limit, counterparty, revocability — issued by a central bank rather than by a protocol consortium, and reachable by every institution in the country. The BCB's stated ambition was to reach up to 60 million Brazilians without credit cards.

What we cannot tell you is how much runs through it. We looked hard for an absolute figure and did not find one. PCMI's BCB-based statistics page carries no Pix Automático adoption data — we checked that page specifically. A web search surfaced a plausible-sounding "about 282 thousand transactions through November 2025"; we went looking for its source and it has none we could find, in the search result's own cited article or anywhere else. We are recording that here rather than quietly dropping it, because a confident invented number produced by a summarising layer is precisely the failure this page is about.

What does exist is growth rates from processors. PagBrasil reported on 26 June 2026 that between Q4 2025 and Q1 2026 its own Pix Automático transactions grew 182%, new users 181%, recurring users 177% and processed revenue 170%, and states plainly these are "dados operacionais iniciais da PagBrasil" for its own ecosystem, not BCB data. The same four percentages appear in trade coverage elsewhere, in a piece written by PagBrasil's own CEO — restatement, not corroboration. A 182% increase on an undisclosed base is a marketing artefact, not a measurement. The honest summary: the mandate primitive exists, is mandatory infrastructure on the payer side, is growing quickly at individual processors, and has no published absolute volume.

The consent surface: JSR and biometrics

The piece an agent actually needs — authorisation without a human app-switch — is Jornada Sem Redirecionamento (JSR), a BCB-designed flow announced to global merchants by dLocal on 2 December 2025. JSR puts biometric approval "right inside the checkout, with no redirects, no banking app, and no extra steps", producing what dLocal calls "a card-on-file-like experience". Note the tense in the announcement itself, because this page's whole business is telling shipped from queued: "The rollout of Pix with Biometrics for merchant integrations is beginning, allowing merchants to initiate the pilot phase." That is a rollout starting and a pilot opening, not a product in general use. No conversion or adoption figures accompanied the announcement, and we found none since.

The biometric story also has a hardware wing that deserves hard discounting. At Autocom 2026, Positivo Tecnologia showed a palm-recognition terminal built on Tencent PalmAI, claiming ~500ms contactless verification "proven across large-scale retail deployments" and 99.9% accuracy in production, settling over Pix. Rollout to Brazilian acquirers was stated for the second half of 2026. Both performance figures are Tencent's own, published on Tencent's own site, with no third-party validation cited anywhere on the page. As of 3 August 2026 this is a trade-show demo with a rollout date, not a deployed checkout — and a fast palm scan authenticates a *human standing at a terminal*, which is the opposite of the problem an agent has.

Who plugs in from outside

Boku received its Payment Institution licence from the Banco Central in April 2025 and announced go-live in May 2026 — thirteen months from licence to first transaction, a useful calibration for anyone modelling entry timelines into Brazilian rails. At launch it supports one-time Pix only; Pix Automático, Pix JSR and tokenised recurring payments are roadmap, not product. Boku's accompanying claims — over 150 million Brazilian users, approximately 70% of the population, more domestic transactions than Visa and Mastercard combined — are marketing copy, and the first is *below* the 162.3 million individuals in the BCB-attributed register as of November 2025, which is the ordinary direction of error when a vendor recycles an old statistic.

The relevant reading for a seller: the newest regulated foreign entrant to the best instant rail in the world went live with the least agent-compatible product on it.

The rest of LatAm: same pattern, earlier in it

Brazil is the extreme case of a regional pattern, not an exception to one. Colombia's Bre-B, launched by Banco de la República in 2025, passed 34 million registered users by 20 April 2026 and took the Central Banking Awards 2026 prize for retail payments infrastructure — a public interoperable instant rail that, against a 2024 Colombian population of 52.9 million on World Bank figures, covers a large share of Colombians within its first year. Widely-cited key and transaction counts for Bre-B could not be fetched from a source that would load for us, so we omit them rather than repeat them. What matters here is the shape: a second Latin American central bank has now built the same kind of public instant rail, and it too has no agentic layer, no mandate credential and no agent-identity scheme.

Mexico is usually placed at the other end of the regional ranking, with SPEI dominant and the consumer-facing CoDi and DiMo layers underused. We could not verify current Mexican adoption figures from a source we were able to fetch, so this page makes no numerical claim about Mexico.

The consequence for the region is that the Santander–Visa pilot's five-country footprint — Argentina, Brazil, Chile, Mexico, Uruguay — ran on card rails in every one of them, including the countries with the best public instant rails on the continent. That is not an accident of vendor choice. It is what happens when the only agent-payment protocols in existence are card-shaped.

So: does agentic payment exist in Brazil?

Three things exist. None of them is what the phrase normally means.

An MCP server for Pix. On 21 May 2026 the Brazilian infrastructure firm Iniciador launched what its coverage calls the first agentic-payments MCP over Pix in Brazil — the claim is national, not global, and we corrected that in review. It is a JSON-RPC-over-HTTP MCP server that lets an LLM agent initiate a Pix payment through Open Finance, settling on Pix, with FIDO2 biometric authorisation bound to a device key. Read the architecture carefully, because it is the finding — and read it against the vendor's own product page, which is more interesting than its press coverage. The device is authorised once: Iniciador advertises "1 authorization — 'n' payments", where the "User authorizes their device once in the agentic flow, and payments go through without ever opening the bank app". What repeats per payment is the *confirmation*, not the enrolment: "User always confirms amount and recipient in Iniciador's interface — not the agent's", with Face ID, Touch ID or equivalent, and "The agent never moves money on its own — the human approves every Pix". So far this is conversational payment initiation with a human in the loop at settlement, not an agent spending on its own account. But the same page also states that the product "Supports immediate Pix and scheduling up to 90 days out — instant settlement at execution, no re-authorization needed". A payment that executes ninety days later without a fresh human act is standing authority, in a narrow form: fixed amount, fixed recipient, fixed date, set by a human at authorisation time. That is closer to a scheduled transfer than to a mandate — there is no spending envelope an agent can draw against, and no delegation to a *named agent* — but it is not the absolute per-transaction constraint an earlier draft of this page asserted, and the correction matters because the whole thesis rests here. Iniciador publishes no transaction volume, client count or launch date; its site says only "Agentic Pix is live". Its 174 million payers and 1 billion enabled accounts describe the Brazilian market, not its own customers.

A card-rail pilot. On 12 March 2026 Santander and Visa completed what they billed as Latin America's first end-to-end AI-agent payments, across Argentina, Brazil, Chile, Mexico and Uruguay on Visa Intelligent Commerce. The Brazilian leg bought chocolates. It was explicitly executed "within Visa and Banco Santander's regulated payment framework and under existing parameters and supervisory standards", with no transaction counts disclosed. Note the rail: the first agentic transaction in Brazil ran on Visa cards, not on Pix. Visa extended its Agentic Ready testing programme to Latin America on 29 April 2026 with 85+ additional organisations; that announcement does not mention Pix or any local rail either.

Conversational bots on WhatsApp. Magie, live since early 2024, does Pix and boletos inside WhatsApp over Open Finance initiation. It is an assistant with a confirm button, not an autonomous buyer. The whole national Open Finance initiation channel it belongs to moved just over R$4.2 billion in the first half of 2025 — against a Pix system that had settled R$28 trillion in the ten months to October 2025. Those two periods are not the same length, so normalise before dividing: annualised, the Open Finance initiation channel runs at about R$8.4bn a year and Pix at about R$33.6tn a year, which puts the programmable, agent-friendly path into Pix at roughly 0.025% of Pix — one fortieth of one percent. Magie's own R$1bn volume and 2% share are company statements relayed without independent verification, and are used here only as an order of magnitude.

The absence, stated plainly

There is no Brazilian agentic payment protocol. Nothing in Brazil corresponds to AP2's mandate credential, x402's wire protocol, or China's APOP and A2P2 (China Shipped Agent Payments First. The 1,000x Volume Gap Does Not Survive the Numbers.). As of 17 July 2026 the Banco Central had not regulated AI-initiated payments at all; a consultant quoted in that reporting expected a public consultation "in the next two to three months", which would place it around September or October 2026. Salesmart S.r.l., which publishes Sinapsi, tested this itself on 3 August 2026 rather than trusting the trade report: we found no BCB public consultation on agent-initiated payments, and the BCB's own 2025-2026 regulatory agenda addresses AI as a supervisory and fraud-risk topic, not as a payment initiator. The identified gaps are the ones every jurisdiction has: no agent-identification standard ("Know Your Agent"), no interoperability rules, no way for a merchant to verify that the agent presenting itself is legitimate.

A vendor with every incentive to say otherwise says half of this. Juspay — an AP2 founding partner selling agentic orchestration into the region — wrote in May 2026 that Brazil "has the lowest friction path to agentic commerce among major emerging markets" but "does not yet have the agent-side ecosystem — the AI platforms, the enterprise integrations, the merchant-facing tooling", and that "the limiting factor will not be payment infrastructure. That problem is already solved." We originally read Juspay as also conceding that no operational agentic standard exists locally. On re-checking, it says no such thing — globally it argues agentic commerce "has crossed from pilot to production". So the honest version is narrower than the one we first wrote: Juspay corroborates the missing *ecosystem*, not the missing *protocol*. The protocol absence rests on our own reading of the sources above.

The demand-side evidence is the weakest link and should be treated accordingly. Visa's June 2026 consumer study reports 76% of Brazilians intending to use AI agents for purchases against 44% of Americans, and 46% versus 23% trusting an agent with payment data. Sample 4,000 across four markets, 1,000 each, measured on simulated shopping journeys, commissioned by the company selling the infrastructure, with field dates undisclosed. Stated intent in a vendor simulation is the single least predictive quantity in this entire wiki.

Our forecast, marked as ours

Everything above is sourced fact or a labelled absence. This section is prediction, and it is Sinapsi's, not a source's.

We expect Brazil to be the first country outside China where autonomous agent payments run at consumer scale on a public, non-card, non-crypto rail, and we expect it to happen by an extension of Pix Automático rather than by importing a Western protocol. The reasoning is structural: the two components that everywhere else must be invented and governed by consortium — a revocable spending mandate with a per-transaction ceiling, and a consent surface that authorises without a redirect — already exist in Brazil as regulated public infrastructure — the mandate primitive compulsory for every account provider on the payer side under Resolução BCB nº 402, with no licensing negotiation required. The missing pieces are agent identity and machine issuance of a mandate: a way for a payer to grant a *named agent* an authorisation, and for a merchant to verify it. That is a rulebook amendment to something that already runs, not a new rail.

Against ourselves, three real objections. Card networks are moving faster in-country than the BCB is — the Santander pilot happened in March 2026 while regulation had not started by July — and incumbency plus global protocol alignment may simply win the agentic use case on cards, leaving Pix as the rail consumers use when they are the ones deciding. The BCB has been notably conservative on Pix limits and security, and delegated machine spending is precisely the risk profile that invites caution rather than speed. And an authorisation surface built around biometrics — JSR, Pix Biometria, palm — is an architecture that assumes a present human body, which is a curious foundation on which to build absent-human payments. Iniciador's design, biometric approval per transaction, is that assumption made concrete, and it may be the ceiling rather than the floor.

The general lesson travels beyond Brazil, and is the same one visible in the earning half of The Earn-Spend Loop: Why Machine Payment Is Half-Built: excellent settlement infrastructure does not produce an agent economy by itself. What is scarce is not the ability to move money in 300 milliseconds. It is a machine-legible statement of *who authorised what, to whom, within which limits* — and the countries with the best rails have not obviously got a head start on that, because it is not a rails problem.

What would falsify this

Four checkable events, none of which had occurred by 3 August 2026.

1. A BCB public consultation on agent-initiated payments. Expected around Q3/Q4 2026 by the specialist cited above; absent from the BCB's published agenda when we checked. Its content — whether it treats an agent as a payment initiator, an identity, or neither — decides most of the forecast above. 2. A published absolute volume for Pix Automático. Any BCB or processor disclosure of transactions rather than growth rates would let the mandate primitive be measured instead of asserted, and would settle whether 182% is impressive or trivial. 3. A Pix authorisation granted to a named agent, with a spending envelope, rather than to a device. Iniciador already schedules ninety days ahead without re-authorisation, so "the human confirms every payment" is not the line. The line is delegation: an authorisation whose beneficiary is an agent identity rather than a fixed recipient-and-amount set by a human at signing time. Whoever ships that — Iniciador, a bank, or the BCB rulebook — turns "excellent rail, no protocol" into "protocol". 4. Agent-initiated volume disclosed on cards in Brazil. Visa and Santander have run a pilot and disclosed no counts. A number would tell us whether the card rail is winning the use case that Pix is architecturally better suited to — the outcome that would most cleanly falsify us.

Until then the accurate sentence about Brazil is neither "agentic payments are here" nor "Brazil is behind". It is that Brazil built, for entirely non-agentic reasons, the rail an agent economy would want, and has not yet been asked to open it to agents. Compare the mirror image in The x402 Market, Counted: 14,766 Listings, 12,741 Calls a Day, One Winner, where a protocol built expressly for agents has the standard and almost none of the volume. Brazil has the volume and none of the standard.