Axon Finance

Posted on Mar 18

Your AI Agent Doesn't Need a Wallet - It Needs a Treasury

#ai #web3 #typescript #blockchain

Most AI agent frameworks tell you to "give your agent a wallet." So you generate a key pair, fund it with ETH for gas and USDC for payments, and hope nothing goes wrong.

Now imagine you have 50 agents. That's 50 wallets, 50 private keys, 50 gas balances to monitor. One compromised key and an agent drains everything it holds.

This is the wrong model.

Agents should sign, not hold

The mental shift: agents don't need wallets. They need permission to request payments from a vault they don't control.

Here's the pattern:

Owner deploys a vault (a smart contract they control)
Owner registers agent public keys with spending limits
Agent signs a payment intent (EIP-712 typed data) when it needs to pay for something
Relayer validates the signature, checks policies, submits on-chain
Vault verifies the agent is authorized and the amount is within limits

The agent never holds funds. Never pays gas. Never has withdrawal access.

What a payment intent looks like

// Agent-side code
import { AxonClient } from '@axonfi/sdk';

const axon = new AxonClient({
  relayerUrl: 'https://relay.axonfi.xyz',
  botPrivateKey: process.env.BOT_KEY,
  vaultAddress: '0x...',
  chainId: 8453, // Base
});

const result = await axon.pay({
  to: '0xvendor...',
  token: '0x833589fCD6eDb6E08f4c7C32D4f71b54bdA02913', // USDC on Base
  amount: '5.00', // 5 USDC
  memo: 'API subscription renewal',
});

console.log(result.txHash);

Under the hood, the SDK signs an EIP-712 PaymentIntent struct and POSTs it to the relayer. The agent's private key never touches funds - it only proves "I am authorized to request this payment."

Spending policies

The vault owner sets per-agent limits:

Per-transaction cap - enforced on-chain, can't be bypassed
Daily/weekly/monthly limits - enforced by the relayer
Destination whitelist - agent can only pay approved addresses
Destination blacklist - blocked addresses (always wins over whitelist)
AI verification threshold - payments above X trigger a 3-agent AI review

All configurable from a dashboard. No code changes, no redeployment.

What if an agent key is compromised?

Worst case: the attacker can sign payment intents. But:

They're capped by maxPerTxAmount (on-chain, immutable per bot)
Daily limits stop runaway spending
Whitelist restricts where funds can go
Owner can pause the bot instantly
The attacker can't withdraw, can't change limits, can't add new bots

Compare this to a compromised wallet: attacker drains everything, no limits, no recourse.

The infrastructure gap

Every agent framework (LangChain, CrewAI, AutoGen, ElizaOS) gives you tool calling, memory, and planning. None of them solve the money problem properly.

"Just use an MPC wallet" still means your agent holds funds. "Use a multisig" means your agent can't transact autonomously. "Use a custodial API" means trusting a third party with your money.

Non-custodial vaults with scoped permissions is the missing piece.

Try it

npm install @axonfi/sdk
# or
pip install axonfi

Open source: github.com/axonfi

Building with AI agents and need payment infrastructure? The TypeScript SDK and Python SDK are on npm/PyPI. LangChain integration available.

DEV Community