The First Bank Built for AI Agents

Every agent commerce announcement so far has described how an AI agent shops. Browse products, compare prices, pick the best option. The shopping part gets all the attention. The money still lives in the human's wallet. Catena is the first company to argue that agents should have their own.

An agent can't open a bank account. It can't sign a contract. It has no legal identity. So when an agent pays for something today, it borrows a human's credit card, or spends from a prepaid balance held by a platform, or calls an API that abstracts away the financial plumbing entirely.

The agent never touches money. It touches someone else's money, through someone else's infrastructure, under someone else's compliance obligations.

On September 18, the OCC granted preliminary conditional approval for Catena Trust Bank, N.A., a national trust bank built specifically for AI agent financial services. Fiduciary custody, investment management, and trust services, all under federal supervision.

Catena was co-founded by Sean Neville and Matt Venables. Neville co-founded Circle and co-created USDC.

What a trust bank charter unlocks

A trust bank is a specific type of institution. It doesn't take deposits like a commercial bank. It holds and manages assets on behalf of others in a fiduciary capacity, meaning it has a legal obligation to act in the beneficiary's interest.

For most fintechs handling money, the baseline regulatory path is a money transmitter license. That lets you move funds from point A to point B. But a money transmitter can't hold assets, manage investments, or provide custody services. And each state requires its own license. 50-plus jurisdictions, each with separate requirements.

A national trust bank charter from the OCC collapses that into one federal framework. One regulator, one set of capital requirements, one compliance standard across all 50 states. The charter also lets the institution do things a money transmitter structurally cannot: hold digital assets in custody, manage them according to encoded mandates, and act as a fiduciary for the asset owner.

For AI agents, that distinction matters. An agent that spends from a prepaid balance on a platform is using someone else's custodial relationship. An agent whose funds sit in a nationally chartered trust bank has its own. The trust bank holds the assets, enforces the spending rules, and answers to a federal regulator.

How agent spending authority works

Catena's architecture gives each AI agent a cryptographic identity, a keypair where the private key lives in a secure enclave, not accessible to the agent itself. A policy contract defines what the agent can do with money: per-transaction limits, cumulative spending caps, approved merchant categories, approved counterparties, time-of-day restrictions.

When an agent initiates a payment, it signs the request with its private key. The trust bank validates three things: the signature is authentic, the transaction satisfies every policy constraint, and the human principal hasn't revoked authorization. If any check fails, the transaction is rejected or escalated to the human for approval.

Corporate card programs work the same way. An employee gets a card, the company sets spending rules, transactions that violate the rules get declined. The difference is enforcement. Corporate card rules are checked after the transaction by expense management software. An agent's spending rules are enforced cryptographically before the money moves.

As Neville put it: "Stablecoins are effectively AI-native money. If an AI actor is making a payment, it doesn't make a lot of sense for it to have a physical credit card. It makes a lot of sense for it to have money that moves at internet speed."

That's where the Circle background becomes relevant. USDC is a stablecoin that settles on-chain. Fast, global, programmable. Catena's trust bank provides the custody and fiduciary layer on top of those rails. The person who built the money is now building the bank account that holds it.

Three architectures for agent money

Three distinct models are emerging for how AI agents access financial services. Each makes different tradeoffs between distribution, control, and regulatory weight.

Stripe's model keeps the agent inside Stripe's infrastructure. The agent pays through Stripe, which handles compliance, fraud, and settlement. The Machine Payments Protocol enables agent-to-agent settlement, and OpenRouter provides model routing.

The advantage is distribution. 88% of the Forbes AI 50 already use Stripe. The tradeoff: the agent has no financial identity of its own.

Skyfire's model gives agents prepaid wallets funded by the agent's principal. The wallets sit at partner banks, not a Skyfire-owned institution. Skyfire raised $9.5 million from investors including a16z's CSX and Coinbase Ventures and focuses on developer integration. The advantage is speed, a simple API for agent payments. The tradeoff: the agent's funds are custodied by someone else's bank, under someone else's charter.

Catena's model gives the agent a regulated financial identity backed by a nationally chartered trust bank. The agent's funds are held in fiduciary custody. Spending rules are enforced cryptographically. The advantage is that the agent has its own financial relationship, supervised by a federal regulator. But a trust bank charter takes time to operationalize, and the infrastructure is heavier than an API call.

	Stripe	Skyfire	Catena
Agent identity	None (uses Stripe's)	Prepaid wallet	Cryptographic keypair
Who holds funds	Stripe	Partner banks	Catena Trust Bank
Fiduciary duty	Merchant-of-record	Partner bank	Trust bank (OCC)
Spending controls	API-level limits	Wallet balance	Policy contract
Regulatory basis	Money transmitter	Partner bank charter	National trust charter

The question underneath all three is who bears the fiduciary responsibility when an agent moves money. Stripe absorbs it into its existing merchant-of-record framework. Skyfire delegates it to partner banks. Catena is building an institution specifically designed to carry it.

What "conditional" means

The OCC's preliminary conditional approval is an endorsement of the application, not a license to operate. Catena must still satisfy a list of conditions before opening, typically including minimum capital funded, board and senior management vetted, BSA/AML compliance program fully operational, and IT security validated.

Based on precedent, the timeline from conditional approval to full operation is 6 to 18 months. Anchorage Digital received its conditional OCC trust charter in January 2021 for crypto custody. Paxos received theirs in April 2021. Both focused on digital asset custody under federal supervision.

Catena is the first OCC trust charter specifically framed around AI agent financial services. The conditional approval means the OCC has decided this category of institution should exist. What remains is proving the institution can operate safely.

$48 million in funding ($18 million seed led by a16z crypto in May 2025, $30 million Series A in May 2026), backing from Circle Ventures and Coinbase Ventures, and a co-founder who already built a $74 billion stablecoin.

The question for Catena isn't whether agents need financial infrastructure. It's whether a federally chartered trust bank is the right shape for it, and whether they can clear the OCC's conditions before someone else builds a faster alternative on lighter regulatory rails.