World Liberty Financial launches AgentPay SDK, an open-source toolkit letting AI agents hold, move, and settle USD1 across EVM chains with built-in policy controls.
AI systems reason well. Paying for things? That is still a problem. World Liberty Financial is trying to close that gap with a new open-source toolkit built specifically for autonomous agents operating with real money.
The project, posted by @worldlibertyfi on X, describes a gap between what AI systems can reason through and what they can actually do financially. Agents need to hold funds, move money, settle obligations, and act within clear spending rules. Payment rails for that kind of work have not existed. Until now, maybe.
When Your Agent Needs to Pay a Contractor
World Liberty Financial, on X, introduced the AgentPay SDK as an open-source, self-custodial toolkit designed from scratch for non-human transactors. It ties into tools developers already use, including Claude Code, Codex, Cursor, Windsurf, Cline, and OpenClaw. The SDK runs locally on a user’s machine. Zero data goes back to WLFI.
USD1, the dollar-denominated stablecoin backing the whole thing, sits at the center of the payment flow. It is pre-configured on both Ethereum and BSC at a fixed contract address. The SDK does a balance check, runs a policy evaluation, signs locally via Unix domain socket, then broadcasts. Private keys never touch the agent or any external service.
That pipeline matters. The signing happens locally through a vault-daemon. Never over the network. It is a small technical detail with significant implications for anyone running agents in production.
The Machine Payments Protocol from Stripe is moving in a similar direction, letting AI agents transact across HTTP-addressable endpoints without human intervention. The infrastructure race for autonomous agent payments is clearly on.
What Happens When Things Break
Three scenarios ship with the SDK’s core design logic. The routine case, where a transfer clears automatically. The insufficient funds case, where the SDK stops, surfaces exactly what’s missing and shows the agent a QR code for mobile top-up. No failed transactions. No wasted gas on a doomed broadcast.
Then there’s the approval threshold case. An agent tries to pay 200 USD1 to a contractor. The per-transaction limit is set at 100. The SDK pauses. Generates an approval request. A human reviews and approves via CLI. Transaction signs and broadcasts.
Routine payments run. Higher-risk ones stop and wait. That is the design logic, stated simply.
World Liberty Financial, in the X post, called this “a recoverable workflow” rather than a failed request. The distinction points to something operators building with AI agents run into constantly. An agent that fails silently is dangerous. One that fails transparently and asks for help is usable.
Bitrefill, 40 Commands, and What’s Next
The SDK ships with over 40 CLI commands covering wallet management, token transfers, policy editing, chain switching, RPC queries, relay configuration, backup, and recovery. Bitrefill is built in, which means agents can buy gift cards, eSIMs, and prepaid products across 12 EVM payment methods without leaving the terminal.
AI agent payment infrastructure is converging fast. Coinbase has already moved AI agents to stablecoin wallets, signaling the same structural push toward agent-native financial rails. The AgentPay SDK takes a more opinionated approach, building policy enforcement and human approval gates directly into the transaction layer.
According to @worldlibertyfi on X, the roadmap extends beyond what shipped today. EIP-3009 meta-transactions are in active development, which would let agents transact without needing native gas tokens. A white paper on AI agent payment security is planned. A formal EIP proposal for policy-aware agent interfaces is also on track.
The plug-in ecosystem is set for third-party extensions that do not require modifying the SDK core. Cross-border payments, remittance, institutional settlement, and DeFi protocol integrations are all listed as targets.
What USD1 is being positioned as, per the @worldlibertyfi X post, is an economic operating layer for autonomous systems. Not another stablecoin narrative built for earlier market conditions. Something built for what money actually does when agents are the ones spending it.
The installation is one line on macOS. The infrastructure, though, is trying to be something that takes longer to build.
