This week, Catena Labs raised $30M from a16z to build an "AI-native bank." x402 Foundation launched operationally under the Linux Foundation to standardize internet-native payments. Visa announced a live proof of concept for AI agent commerce.
Coverage treated these as competitive moves. They're not.
The AI agent payments market isn't consolidating into a winner-take-all race. It's layering into a stack. And most teams building in this space — including investors writing the checks — haven't mapped the layers clearly.
Here's the actual structure.
Layer 1: Protocol
What it does: Defines the standard for how agents request and authorize payments.
Who's building it: x402 Foundation (Coinbase, Stripe, others), Stripe's Micropayment Protocol (MPP).
Example: x402 extends HTTP with a 402 Payment Required status code. An API returns 402, the agent's wallet authorizes payment, the API grants access. All done in milliseconds, programmatically.
This is infrastructure. It's not a product users interact with. It's the plumbing.
Layer 2: Banking & Custody
What it does: Holds funds, manages compliance, provides trust infrastructure.
Who's building it: Catena Labs (filing for a national trust bank charter), custodians like Fireblocks.
Example: An enterprise deploys 1,000 agents. Those agents need wallets funded with USDC. Catena provides the bank accounts, regulatory compliance, and institutional custody.
This is the backend for agent finance. Enterprises need it. Developers don't interact with it directly.
Layer 3: Gateway
What it does: Translates agent intent into executable payments. Manages wallets, policies, authorization logic, and settlement.
Who's building it: AgentWallex, Skyfire.
Example: An agent needs to pay an API. The gateway checks policy rules (spending limit, recipient allowlist), generates an MPC signature, routes payment through the appropriate rail (x402 micropayment or stablecoin transfer), and logs the transaction. All in <150ms.
This is the integration layer. Developers building AI agents interact with this via SDK.
Layer 4: Legacy Adaptation
What it does: Lets agents use existing payment rails built for humans.
Who's building it: Visa (virtual cards for agents), Mastercard, PayPal.
Example: Agent gets a virtual card number. Uses it to pay a non-crypto API that only accepts cards. Visa processes it like a normal card transaction.
This works, but it's retrofitting. Cards weren't designed for autonomous, high-frequency, micropayment use cases.
Why This Matters
Most coverage treats Catena, x402, and AgentWallex as competitors. They're not.
- Catena is building the bank.
- x402 is building the protocol.
- AgentWallex is building the gateway.
These are complementary, not competitive.
An agent developer doesn't choose between them. They use all three:
- Their enterprise customer's funds sit in a Catena trust account (Layer 2).
- The agent pays an API using x402 (Layer 1).
- AgentWallex provides the wallet, policy engine, and authorization logic (Layer 3).
Where AgentWallex Fits
We're the gateway layer. We sit between agent intent and payment execution.
Our job:
- Give agents MPC-secured wallets (via Paratro) so they can pay autonomously without exposing private keys.
- Enforce policy rules (spending caps, recipient allowlists, rate limits) without requiring manual approvals.
- Support x402 micropayments natively — authorize and settle in <150ms.
- Route payments to the appropriate rail (stablecoin transfer, x402 call, future: virtual card for Web2 APIs).
We don't compete with x402. We support it.
We don't compete with Catena. We integrate with custody providers.
We don't compete with Visa. We route to them when an agent needs to pay a card-only merchant.
The Market Is Real
$73M in AI agent payments processed via crypto rails over the past 12 months. That's not speculative. That's agents actually paying for compute, API calls, and services.
Binance integrated x402 this week. Agents can now buy services without human approval.
Robinhood, OKX, and Injective are all shipping AI agent trading features.
The infrastructure is no longer hypothetical.
What Developers Need
If you're building AI agents with LangChain, CrewAI, or AutoGPT, you don't need a bank charter. You don't need to implement x402 from scratch. You don't need to issue virtual cards.
You need a gateway.
You need an SDK that gives your agent a wallet, lets it authorize payments in code, and handles the complexity of routing, policy enforcement, and settlement.
That's what we built.
Payer SDK: Your agent gets an MPC wallet. Authorize payments in <150ms. No private keys exposed.
Merchant SDK: Your API accepts agent payments. Charge per call, per result, per token. x402 native.
Policy Engine: Set per-agent limits, allowlists, time-based rules. No manual approvals. Full audit trail.
Stablecoin-first (USDC on Base for MVP). Expanding to multi-chain, multi-rail.
The Stack Is Layering
Everyone building AI agent payments is solving a real problem. But they're solving different problems at different layers.
Catena is solving institutional custody.
x402 is solving protocol standardization.
Visa is solving legacy compatibility.
AgentWallex is solving developer integration.
The market structure isn't winner-take-all. It's modular.
And we're the gateway layer — the point where agent intent becomes actual payment, regardless of what's beneath it.
Sandbox live at app.agentwallex.com. 3,600+ teams on the waitlist.
Follow & Try AgentWallex
- 🌐 Website: app.agentwallex.com
- 🚀 Sandbox (free): app.agentwallex.com
- 📖 Docs: docs.agentwallex.com
- 📲 Telegram: t.me/AgentWallexOfficial
- 🐦 X / Twitter: x.com/AgentWallex
- 🦋 Bluesky: bsky.app/profile/agentwallex.bsky.social
- 💻 Dev.to: dev.to/agentwallex
- 📝 Hashnode: agentwallex.hashnode.dev
Top comments (1)
I particularly appreciated the breakdown of the AI agent payments stack into four distinct layers, which clarifies the complementary nature of various players in the space, such as Catena Labs, x402 Foundation, and AgentWallex. The example of an agent developer using all three - Catena for trust accounts, x402 for payment protocol, and AgentWallex for the gateway - effectively illustrates how these layers work together seamlessly. This layered approach reminds me of the evolution of the web, where different layers of infrastructure and standards enabled widespread adoption. As the AI agent payments market continues to grow, I'm curious to see how these layers will continue to develop and intersect with existing financial systems, such as traditional banking and card networks, to enable more efficient and secure transactions.