An open settlement primitive for agents, and the autonomous treasury agent that's already its first live user.
Treasury payments run on trust, not proof.
No safe way to leave an agent alone
Either a human approves every transfer, or the agent holds keys that, if they leak, empty the whole treasury.
No provable record
A wrong payment is a support ticket and a spreadsheet, not a transaction anyone outside the company can check.
No settlement fast enough to watch closely
If checking a balance costs almost as much as the payment, nobody checks it often enough to catch problems early.
A ledger-writing machine, not a script with a cron job.
Arcurrent reads what's owed, decides from the real vault balance, the due date, and a reserve floor, and pays from an on-chain vault it can never withdraw from and can only spend inside limits its owner set. Every payment is a mandate on MandateEscrow, an open primitive any address can fund, fulfill, or read.
Nothing here is simulated. Every figure on the next slide is a real transaction hash on Arc mainnet, not a mock.
Two real settlements, five real transactions, on Arc mainnet.
Captured 2026‑09‑22, real mainnet USDC. Current counts are on the live dashboard and Mandates table.
Two obligations, $0.05 and $0.03 USDC, settled by creating and releasing mandates on MandateEscrow · not a pre-funded pool, straight from the treasury wallet's own balance.
Fulfiller reputation after both: 2 completed, 0 refunded, $0.08 volume settled · read live from the contract, not this project's database.
Bonus proof, captured 2026‑09‑23: the one capability the two settlements above don't show, since both paid a single destination.
AgentVault proof, captured 2026‑09‑28: the payment path this deck actually describes, not the direct-wallet flow above. The vault is the funder of record, not the operator wallet.
Six real pieces, one decision loop.
Circle wallet as operator
The agent signs from a Circle-custodied wallet that holds only gas. No private key lives in this codebase, and nothing worth stealing lives in the wallet.
AgentVault
The treasury sits in a contract. The owner sets a per-payment cap, a daily cap that refills continuously, a payee allowlist and a pause switch. The agent can only pay inside them and can never withdraw, so leaked credentials are bounded by numbers the owner chose.
MandateEscrow
A permissionless contract on Arc mainnet: fund a mandate for anyone, prove fulfillment, release atomically, split payouts, refund on deadline. Reputation updates on-chain. Not project-owned · this agent is just its first live caller.
App Kit · Bridge (CCTP)
Circle's App Kit SDK, kit.bridge(): cross-chain USDC top-ups when a payment would breach the reserve floor. Proven on testnet; disabled on mainnet until a Live-environment source wallet is funded.
x402 nanopayments
The agent pays a sub-cent fee for the live FX rate it needs before acting on a foreign bill.
SDK and MCP server
A typed client and an MCP server let any project or AI agent use MandateEscrow, or pay through a vault, with spend caps and simulate-first errors. In vault mode the only way to pay is one bounded tool.
This is the pattern Circle's Agent Stack formalizes: wallets, an onchain program, cross-chain liquidity and a machine-payable service, wired into one loop. Built directly against those primitives rather than a starter kit, since the loop itself is the product.
Arcurrent: current, twice.
Gas paid in USDC, the same money moving through the system, not a second token to keep topped up. Settlement under a second, so a balance check every few minutes doesn't cost more than the payments it's protecting.
That's the whole bet: money that moves as fast as the decision to move it.
The money already moved on-chain. The agent didn't.
USDC already sits in real company treasuries as real settlement money. What's been missing isn't the rail, it's an agent trustworthy enough to be handed the keys and left alone.
Arc is the first L1 built specifically so that agent doesn't need a human chaperone: USDC-native gas, sub-second finality, and a Circle stack built for custody without a human touching a key.
What's real, what's gated.
- realLive on Arc mainnet: Circle Live-environment wallet as operator, AgentVault, MandateEscrow (source verified), x402 nanopayments, autonomous cron, webhook-confirmed settlement end to end, a public SDK and MCP server
- unauditedThe contracts are self-reviewed, not professionally audited: static analysis, fuzz and mutation-checked tests, and a real-chain rehearsal. The vault holds a small balance for that reason.
- gatedStableFX: RFQ-only access, no self-serve signup yet. Non-USDC obligations are correctly flagged, not yet auto-settled. CCTP liquidity top-up currently disabled on mainnet pending a funded source chain.
A primitive, not another single-purpose demo.
Relevance to Arc
Impossible on a chain with volatile gas or probabilistic finality: USDC-native gas and sub-second settlement are what make per-obligation mandates economical at all.
Technical credibility
Real contracts, real mainnet deployment, a real autonomous caller · not a testnet-only illustration. The vault's cap maths is fuzz-tested and mutation-checked, and every rule was attacked on a live chain.
Depth
Solves trust and accountability, atomic multi-party settlement, and portable reputation in one primitive, rather than one narrow feature.
Promise
Open by design: any other agent or hackathon builder can fund, fulfill, or read a mandate today, no permission needed.
Where this goes from here.
What we're looking for.
The microgrant
Non-dilutive fuel to keep building the primitive out, and to pay for an independent review of the contracts.
StableFX access
The one piece that's flagged, not faked. This closes the last gap between "correctly identifies" and "fully autonomous."
Introductions
To a real treasury with genuine recurring obligations to pilot against, and to other Arc builders whose agents could be MandateEscrow's second caller.