硅谷Fintech观察⑧｜谁在为 AI Agent 建金融系统：一张两层地图，和它没画的三格Silicon Valley Fintech Watch ⑧ | Who Is Building the Financial System for AI Agents: A Two-Layer Map, and the Three Cells It Left Blank

Silicon Valley Fintech Watch series · Part Eight · First published on klay-wang.com, August 27, 2026

In brief. A map of the agentic finance stack has been making the rounds: four cells on top for what agents can do with money, six gates below for the questions a system must answer before any dollar moves. The frame is good. But a map is worth only as much as what sits behind each logo: who wrote the checks, who holds the charter, what position it occupies. I went through all two dozen names and checked funding and regulatory records. Three findings turned out to be worth more than the map itself. The incumbents have filled the protocol columns and startups survive only at the two ends. Agentic banking, billed as the next big wave, has exactly two institutions that have walked the charter path all the way. And three cells are missing from the map entirely. The last piece was about how companies pay the models. This one lays out the whole table: which layers of money are being rebuilt for agents.

Before we start

A note on sources and method.

The starting material is an image and its accompanying text circulating on LinkedIn, titled Who is building the financial system for AI agents. The author splits the market into two layers: a front layer of products and services covering what agents can do with money, and a back layer of infrastructure and trust covering six questions the system must answer before an agent moves money. The text opens by sizing the market at three to five trillion dollars within a few years.

I am not using that number. It carries no source and no derivation, and it is the classic top-down figure extrapolated from AI GDP. The house rule in this series is that market size gets built up from billable units: how many enterprise customers, how much each pays, how many basis points get taken. If you cannot build it up, do not cite it. So this piece does not talk trillions. It talks about what can be verified: how much each company in each cell has raised, from whom, under what charter, and where it is stuck.

The method is unglamorous. Two dozen names, each checked against public funding records, regulatory filings and product announcements, with every figure listed at the end. Three of my own impressions changed along the way. I had assumed the protocol layer still had room for startups, and found that nearly every startup in that column is funded by incumbent strategic arms. I had assumed agentic banking was still a vision, and found it already has a shipped product line, sitting on a startlingly small denominator of chartered institutions. And I had assumed the map was complete, and found that what it leaves out is worth more than what it draws. Those three revisions are where the judgments below came from. First, read the map.

1. Reading the map: the top layer spends, the bottom layer governs

The two-layer structure is the most valuable thing about this map, so it is worth restating faithfully.

The front layer is products and services, answering what agents can do with money, in four cells: agentic commerce (agents discover, compare and buy), agentic payments (agents initiate and execute payments for people or businesses), agentic fintech (fintechs expose existing products to agents), and agentic banking (the regulated institution itself is designed to serve agents directly).

The back layer is infrastructure and trust, and it is the sequence of gates a dollar passes before an agent can move it: identity, who is the agent; authorization, who gave it authority; permissions, what exactly is it allowed to do; runtime, where those rules are enforced while it operates; wallets, where funds and signing capability are held; and payment rails, how value actually moves.

That ordering is not arbitrary. It is a chain of accountability: establish the actor, then the delegation, then the boundary, then the execution environment, then custody, and only then the transfer itself. In the existing payment system those six jobs are absorbed by bank onboarding, card network rules and the courts, and the user never sees them. Agents expose all six again, because every link in the old system assumes a human operator: a human logs in, a human clicks confirm, a human bears the dispute. Swap the operator for software and all six gates have to be rebuilt.

One thing that is easy to misread. The circulating image merges authorization and permissions into a single column; the text version keeps them apart. Apart is correct. Authorization asks which person or company the power came from. Permissions asks, once it has the power, how much it can spend, on what, and to whom. Stripe's ACP and SPT belong to the second: a scoped payment token with per-transaction limits and named merchants answers what it can do, not who authorized it. That distinction matters again in section three.

That is the map. What follows is the actual work: verify each cell, then say three things the map cannot show you.

2. The top four cells are a ladder of distance from fiduciary duty, not four parallel lanes

Drawing four cells side by side invites you to read them as four parallel races. Having checked the players in each, my conclusion is the opposite: these are not parallel, they are sequential, and the axis is distance from fiduciary duty. Each rung moves you closer to being answerable for customer money yourself, and the character of the revenue changes with it.

Rung one, agentic commerce: no license on the money, just the shelf. Shopify is the exemplar. It opened its agentic commerce infrastructure to developers through UCP and the Catalog API: UCP is an open standard developed jointly by Shopify and Google for how agents transact with merchants, and the Catalog API is its discovery layer, turning products from millions of merchants into structured, queryable data for agents. The more telling change is in access: this used to require approval, and now it is self-serve, with developers registering an agent profile and calling a public endpoint. Turning an approval process into self-serve is itself a statement of position. Shopify wants agent traffic to be as large as possible, collects rent at the shelf, and takes no liability at the money layer. Google and PayPal play the same logic in this cell: entry point and conversion, no regulatory exposure.

Rung two, agentic payments: the rail owners came down themselves. Visa published its Trusted Agent Protocol in October 2025. Mastercard extended Agent Pay to machine-to-machine scenarios in June 2026. Stripe has MPP and ACP/SPT. All three rail owners are personally building agent credentials and controls, with none sitting it out, which section three takes up in detail. The more informative name in this cell is the unfamiliar one, Natural: founded August 2025, a $9.8 million seed in October, a $30 million Series A in July 2026 at a $150 million post-money valuation led by Forerunner's Kirsten Green, more than $40 million raised in total. Its product list shows where a startup can still cut into the rail layer: thirteen planned products spanning ledgering, money movement, fraud and compliance, six of them already live, including agent wallets carrying FDIC insurance through a partner bank sweep arrangement, and vaults that take money in but never let an agent move it out alone. It is not trying to build another card network; it is building a native account and ledger system for agents. Rail incumbents write the protocols, startups write the ledger. That is the current division of labour on rung two.

Rung three, agentic fintech: the front end is yours, the balance sheet is someone else's. All four names here have verifiable records. Ramp raised $750 million in the second quarter of 2026 at a $44 billion valuation. Mercury raised a $200 million Series D at $5.2 billion. Slash raised $100 million at $1.4 billion, a new unicorn. Meow has raised roughly $33 million in total, with Tiger Global, QED and Lux on the cap table. They share one structure: they own the software and the customer relationship and depend on a chartered bank underneath. And they are doing exactly what the map's author describes. The most aggressive move came from the smallest of them: in April 2026 Meow announced what it called the industry's first agentic platform letting AI agents open and manage business bank accounts on a user's behalf. Agents opening bank accounts went from concept to product announcement in April 2026.

Rung four, agentic banking: the regulated institution recognizes the agent. The map's author calls this a very small category. How small can be quantified with a list. Among crypto-oriented applicants for a US national trust charter, the OCC has granted conditional approval to nine: Circle, Ripple, BitGo, Fidelity, Paxos, Protego, Bridge, Crypto.com and Coinbase. Of those nine, exactly one, Circle, has completed final approval, in July 2026. Add Anchorage, chartered back in 2021 and operating since, and the number of institutions that have walked this charter path all the way is two. That is the denominator for rung four.

But this cell changed character from the second quarter of 2026 onward: it went from slogan to product. Anchorage launched a platform called Agentic Banking in May 2026, in partnership with Google Cloud. The mechanics are worth reading closely. An enterprise funds and governs its own agents; when an agent reaches the point of transaction, the platform first enforces corporate spending policy, Know Your Agent identity standards and real-time compliance checks, and only then settles across stablecoins, fiat rails or tokenized credentials. Note that Know Your Agent has moved from industry jargon into the product language of a chartered bank. Anchorage has also disclosed a pipeline of up to twenty large institutions waiting to issue stablecoins through it.

Two newer names belong in this rung, pointing in opposite directions. Catena Labs, founded by Circle co-founder Sean Neville, has applied to the OCC for a national trust charter; the application has been accepted and is not approved. It has raised $48 million in total, including a $30 million Series A announced May 20, 2026, co-led by a16z crypto and Acrew with QED, General Catalyst and Coinbase Ventures among the participants. Its bet is to build custodial accounts for agents themselves, writing authorization and audit into the legal duties of a regulated institution. The other is Erebor, a national bank that opened in February 2026, with deposits growing from roughly $1.1 billion at the end of March to roughly $4.6 billion at the end of July, having raised $350 million in December 2025 at a $4.35 billion post-money valuation led by Lux; the Financial Times reported on August 11, 2026 that it was in talks to raise $1.5 billion at an $8 billion pre-money valuation (about $9.5 billion post-money), with no closing announced as of publication. Here is a distinction the map does not draw: Erebor is a bank that opens accounts for companies in the AI industry, and its customers are human enterprises; what Catena and Anchorage's Agentic Banking are building is a world where the agent itself is the party being served. The former is a traditional bank reaching a new customer segment. The latter is what rung four actually means. Conflating them will misjudge the size of this category by an order of magnitude.

That is the ladder. To summarize how revenue changes with it: rung one collects shelf rent, rung two collects protocol and network fees, rung three earns software subscriptions and a share of spread, and rung four is paid for accepting fiduciary duty. The higher you climb, the more the revenue looks like rent, the more expensive the ticket, and the heavier the charter. That is the real structure of the front layer.

3. Naming the back layer cell by cell: power is written in the cap tables

The back layer is where this map carries the most information. Here are the players, their core strengths and their funding records, followed by two observations that come out of the table.

Observation one: the authorization and permissions columns contain almost no startups. Column 02 is Google and Mastercard. Column 03 is Stripe. The one startup, nekuda, has raised a $5 million seed, and its cap table contains Amex Ventures and Visa Ventures. The same pattern repeats in identity: Skyfire's $8.5 million seed includes Circle, Gemini and Ripple, with Coinbase Ventures topping up later to $9.5 million in total. Read those lists together and the conclusion is blunt: the startups in the protocol layer are funded mainly by the strategic arms of the rail incumbents. The incumbents publish their own protocols while placing small checks in every startup that might grow into a competing standard. That is not the distribution of financial investment; it is the distribution of defensive positioning. For founders the signal is harsh but clear: build an independent company in authorization or permissions and the best ending your funding story has is being bought or absorbed by a rail owner. There is no precedent on this map for growing into a standalone platform there.

Observation two: the wallet column has the highest crypto concentration on the board, and the sturdiest funding among newly formed companies across the six gates. (Persona in the identity column has raised more, but it is a seven-year-old KYC business changing flags rather than a new entrant, and it gets its own note below.) Crossmint's $23.6 million was led by Ribbit with Franklin Templeton participating. Turnkey has raised more than $65 million, with Bain Capital Crypto leading a $30 million Series B alongside Sequoia, Circle Ventures joining in a strategic round in May 2026, and Stripe's Bridge and Polymarket among its customers. The reason is not complicated. Where funds and signing keys live is a question the crypto industry has spent five years and real money on: key management, multi-party computation, programmable custody. Those are precisely the custody primitives an agent needs. This is one of the few positions on the map where a startup can stand on accumulated technology rather than a charter or a distribution channel. Even the Anchorage and Google Cloud partnership has key management at its center.

Observation three: incumbents have started changing flags. This signal runs across columns, and it shows up in Persona. The company spent seven years doing KYC and ran more than three hundred million verifications in 2024. When it raised a Series D in April 2025, it titled its own press release around building the verified identity layer for an agentic AI world. The $200 million led by Founders Fund and Ribbit is, in effect, an established identity player publicly changing flags. When incumbents start rewriting their funding stories around agents, the capital in that layer has already begun to take sides.

4. Three judgments

The verification is done, the map reread, the table built. Here are my own three judgments, each with its evidence and its falsification condition.

Judgment one: there is no independent startup position in the protocol cells of the back layer. The openings are at the two ends, the regulated asset end and the vertical packaging end.

The evidence is the structure of the table in section three: authorization and permissions are entirely incumbent protocols, and the sole startup there is funded by incumbent strategic arms, while the wallet column has the sturdiest funding and the most real customers. The two ends mean this: go down into the regulated asset layer, or go up into vertical packaging.

One self-correction is needed here, or the judgment reads too optimistically. The regulated asset end is being filled by incumbents too. Of the nine OCC conditional approvals listed in section two, eight are public companies or industry giants, and Bridge is itself a Stripe subsidiary. So the accurate version is that startups are being squeezed on the charter path as well, and the one downhill position not yet taken is the narrow one: signing and custody, work that runs on accumulated technology and that incumbents cannot be bothered to build themselves. That is where Turnkey and Crossmint stand, while Catena's charter route means racing eight giants on the same track. The uphill position is the wider one: building a native ledger for agents the way Natural does, or turning agent account opening into a product the way Meow does, work the incumbents' own org charts are too rigid to turn toward. The protocol layer in between, how credentials get issued, how intent gets expressed, how authority gets passed, is the battleground of Visa, Mastercard, Google and Stripe. They have the installed networks, the distribution and every incentive to give the protocol away for free. There is no defensible position there for a startup.

Falsification: if before the end of February 2028 an independent startup in authorization or permissions closes a Series B or later led by a purely financial investor, with at least two large financial institutions using it as their primary protocol, this judgment is void.

Judgment two: agentic banking just turned from slogan into product, the ticket to this category is a charter, and the denominator on that charter is currently two.

Three pieces of evidence. First, Anchorage has made Agentic Banking a real product line, with policy enforcement, Know Your Agent, real-time compliance and cross-rail settlement. That is the first time a chartered institution has designed a product treating the agent as the account party. Second, the regulatory denominator is checkable: nine conditional approvals from the OCC, one completed (Circle), plus one operating incumbent (Anchorage), two in total, and of those two only Anchorage is actually building products that treat the agent as the account party. Third, the private market has already begun pricing that ticket: Catena has raised $48 million while holding nothing but an accepted application and no disclosed revenue, from a16z crypto, Acrew, QED, General Catalyst and Coinbase Ventures. Valuation for companies like this is fundamentally a charter option: the bet is not on today's cash flow but on the moment the approval lands and fiduciary duty converts into revenue.

A risk note attaches to this judgment: options can expire unexercised. An accepted OCC application is not an approval, and capital requirements, governance conditions and timelines can all shift. If the approval is slow, incumbents doing the same work, Anchorage among them, will consume the window. Anyone treating these companies as a sure thing is buying a regulatory calendar.

Falsification: by the end of 2027, if three or more institutions without their own charter (working through partner banks or state licenses) have built scaled products treating the agent as the account party and disclosed customer numbers, while the chartered path still has only one or two players doing so, then the ticket is a charter claim is void and the correct version becomes the ticket is a partner bank relationship. This is not a hypothetical: Natural, in section two, uses a partner bank to put FDIC insurance behind agent wallets, and Meow lets agents open business accounts, and neither holds its own charter. My current view is that they cannot capture the fee that comes with fiduciary duty and are left selling software, but that view may be wrong.

Judgment three: the real openings are not on this map. Three cells are missing: before, after, and asset management.

Before means the rehearsal and evaluation an agent needs before it touches real money. That cell has started to fill. One day before this piece was published, on August 26, 2026, a company called Arga Labs announced a $10 million seed led by General Catalyst with BoxGroup, Emergence, Gradient and SV Angel participating. It builds full-scale digital twins of enterprise software, cloning systems like Salesforce and Workday with their permission models and webhooks intact, so agents can make their mistakes in a copy first. One founder came from Amazon, the other from Stripe and Goldman Sachs. Once execution costs collapse, the scarce thing becomes the willingness to let an agent touch a live system, and rehearsal is the first station where willingness becomes procedure.

After means audit, disputes and insurance, and that cell is still empty. It may be the most valuable empty cell on the map, and the reason comes from the payments industry's own history: the power of a card network has never been anchored in transfer speed, it is anchored in dispute rules. Whoever defines refunds, allocates liability and absorbs the loss owns the network. Agentic payments mix irreversible stablecoin settlement, revocable ACH and card rails with mature dispute regimes into one interface, and allocating responsibility after something goes wrong is far harder than checking authority before it does. All six gates on the map are before-and-during gates. Not one governs after. The empty cell may mean the real toll position is still unoccupied. It may equally mean no independent player will ever grow there, because disputes and loss allocation were always going to be split between card network rules and the courts. I cannot yet tell these two apart, which is why a falsification condition follows.

Asset management means agents managing money, not just spending it. All six questions on this map are about outflow: what to buy, whom to pay, how to settle. Not one asks how idle corporate cash gets swept by an agent into a money market fund, how a portfolio gets rebalanced by an agent, how a redemption gets triggered. Spending is a liability-side activity; managing is an asset-side one, and the asset-side regulatory stack (advisers, custodians, funds) is deeper than the payments stack and far less agent-enabled. When a player finally shows up in this cell, that is worth watching on its own.

A falsifiable version: by the end of 2027, if maps of this kind still show no after cell (a dedicated player in disputes, audit or agent insurance), and no insurer has disclosed an agent-related underwriting product in its reported figures, then the after half of judgment three is void and the right reading will be that this cell was always going to be split between the card networks and the courts.

5. What this map will not tell you

Before the close, an argument against my own piece. There are five places where this article and the map behind it deserve suspicion.

First, a logo on the board is not volume. Most of the protocols here were announced in 2025 or 2026: Google AP2 in September 2025, Visa TAP in October 2025, Stripe MPP in March 2026, and Mastercard's Agent Pay extension for machines in June 2026, all within the past year. I have found no disclosure of production transaction volume from any of them. The infrastructure-first narrative has another reading: incumbents are occupying standard positions before demand is confirmed, and the announcement is itself a competitive act. Plenty of enterprise agent payments remain at the stage of drafting a payment for a human to approve. Six gates built and nobody walking through them is a possible ending too.

Second, the sample is someone else's, not a market census. This map carries two or three names per cell, chosen by its author. So a cell with no startups in it first means the map did not draw one, and only second might mean the market has none. Every claim in this piece about who is missing from a layer should be discounted accordingly.

Third, the standards war is not over. UCP, AP2, MPP, x402 and TAP come from five different interested parties, overlapping and competing. Payment standards historically take a decade to converge, and some die on the way. This cell-by-cell naming is a cross-section of August 2026, not a final roster.

Fourth, the three-to-five-trillion market size is not used once in this piece as evidence. To repeat: that number has no source. Figures of that kind exist to prop up fundraising decks, not analysis. The verifiable revenue in this market today is what the Naturals sell to enterprises, the custody and fiduciary fees at the Anchorages, and the transaction take at the agent wallets. Build up from there and you get the real size of this sector.

Fifth, the limits of this piece itself: it relies entirely on public information and cannot see undisclosed internal progress; funding and valuation figures come from different points in time and should be read as orders of magnitude, not precise comparisons; and the survivors left records while the attempts that died before seed left none, so the real mortality rate is necessarily higher than the visible sample.

The close

Three sentences.

One: the front layer is a ladder of regulatory depth, not four parallel lanes. From shelf to rail to borrowed charter to owned charter, each rung makes the revenue more like rent, the ticket more expensive. To judge what an agentic fintech is worth, first see which rung it stands on, then whether it has a path upward.

Two: you do not have to guess where the power sits in the back layer, it is written in the cap tables. The protocol layer belongs to incumbents, and whoever's strategic arm is on the board is defending that layer. The asset layer belongs to charters and signing keys, and that is where startups can stand on genuine capability. When you look at a deal in this space, read the investor list before you listen to the story. The list is more honest than the story.

Three: the real openings are off the map. Rehearsal has just seen its first entrant, accountability and asset management are still blank, and the history of payments says the toll position has never been at the execution layer; it sits at the layer that answers who is responsible when something breaks. When the next map like this gets drawn, watch which names land in those three cells. That is probably where the companies worth remembering from this cycle will come from.

(Sourcing. The three-to-five-trillion market size is the circulating text's own figure, carries no source, and is not relied on here. Shopify's UCP, developed jointly with Google, as an open standard, the Catalog API as its discovery layer, and the shift from approval-based to self-serve access come from Shopify's developer documentation and its Spring 2026 Edition announcement. Visa's Trusted Agent Protocol was published in October 2025; Mastercard's Agent Pay first launched in April 2025 with the Agent Pay for Machines extension in June 2026; Stripe MPP and ACP/SPT from Stripe's own materials; Stripe's roughly $159 billion valuation is the February 2026 employee tender basis. Natural: $9.8 million seed in October 2025 (co-led by Abstract and Human Capital), $30 million Series A in July 2026 at $150 million post-money led by Forerunner's Kirsten Green, more than $40 million raised in total, product set including agent wallets carrying FDIC insurance through a partner bank sweep arrangement, per Axios, TechCrunch and company releases. Ramp's $750 million at $44 billion, Mercury's $200 million Series D at $5.2 billion and Slash's $100 million at $1.4 billion are second-quarter 2026 public reporting; Meow's roughly $33 million total and its April 2026 agent account-opening announcement come from company releases and public reporting. Anchorage Digital's Agentic Banking launch, the Google Cloud partnership, the Know Your Agent and policy-enforcement description and the pipeline of up to twenty institutions come from company announcements and May 2026 CoinDesk reporting; its 2021 OCC charter is a matter of public record. The OCC's nine conditional approvals for crypto-oriented national trust charters (Circle, Ripple, BitGo, Fidelity, Paxos, Protego, Bridge, Crypto.com, Coinbase) and the single completed approval, Circle, in July 2026, reflect the public OCC record as compiled in late August 2026. Catena Labs: $48 million raised in total, a $30 million Series A announced May 20, 2026, co-led by a16z crypto and Acrew with Breyer, General Catalyst, QED, Oak HC/FT, Fin Capital, Coinbase Ventures and IDG participating; its OCC national trust charter application is accepted and not approved, and the public application states it will not engage in ordinary lending to the public, per company and investor announcements and the OCC public filing. Erebor: OCC conditional approval in October 2025, FDIC deposit insurance in December 2025, opened February 2026; deposits of roughly $1.1 billion at the end of March per regulatory disclosure, roughly $4.05 billion at the end of the second quarter per Bloomberg-sourced reporting and roughly $4.6 billion at the end of July per regulatory disclosure; the $350 million raise at a $4.35 billion post-money led by Lux in December 2025 is public reporting; the talks to raise $1.5 billion at $8 billion pre-money (about $9.5 billion post-money) were reported by the Financial Times on August 11, 2026, with no closing announced as of publication. Skyfire: an $8.5 million seed in August 2024 topped up to $9.5 million in total, with its investor list, from August 2024 and subsequent public reporting. Persona: $200 million Series D at $2 billion led by Founders Fund and Ribbit, more than 300 million verifications in 2024, from April 2025 company and press announcements. nekuda: $5 million seed led by Madrona with Amex Ventures and Visa Ventures, from the May 2025 company release. Crossmint: $23.6 million led by Ribbit with Franklin Templeton and others, from March 2025 company and press announcements. Turnkey: a $30 million Series B in June 2025 led by Bain Capital Crypto with Sequoia participating, a $12.5 million strategic round in May 2026 including Circle Ventures, more than $65 million in total, customers including Bridge and Polymarket, from June 2025 and later public reporting. Arga Labs: $10 million seed led by General Catalyst with BoxGroup, Emergence, Gradient and SV Angel, and founder backgrounds, from TechCrunch reporting on August 26, 2026 and company materials. All positioning and judgment passages are the author's analysis and are not disclosures by any company.)