Agents need to pay each other

AI is moving from assistants to agents — software that doesn’t just answer questions but takes actions, makes decisions, and operates continuously. There will be billions of agents, each acting on behalf of a person, a business, or another agent.

They’ll eventually need to transact with each other.

An agent managing a DeFi portfolio might need to rebalance dozens of positions per hour. An agent running inference jobs needs to pay GPU providers per task. An agent coordinating with other agents needs to settle micropayments in real time.

Current infrastructure wasn’t designed for software that acts autonomously. Every crypto wallet expects a human to sign each transaction, banks expect a legal entity on file, and payment APIs expect OAuth tokens a human set up by hand.

An AI agent making financial decisions hits a wall fast — stopping to ask a human to click “approve” on every action doesn’t scale.

A few primitives are emerging to close this: x402 for HTTP-native micropayments, ERC-8004 for on-chain agent identity, and policy oracles that enforce spending limits without human sign-off on each action. ChainGPT’s AIVM is building a Layer-1 specifically around these primitives.

I wrote a primer on how the pieces fit together: Web3 & AIVM Primer.