← Back to blog

“Trapped in a Box”:
Why AI Agents
Need Their Own Money

Marc Andreessen's argument cuts to the heart of what's missing in the current AI stack: without the ability to transact independently, agents can give advice but they can't take action. Here's what it takes to fix that.

5 min read
WITHOUT FINANCIAL IDENTITY pay invoice buy credits settle escrow sweep yield economically inert Proco WITH ECONOMIC IDENTITY pay invoice buy credits settle escrow sweep yield economically autonomous

Marc Andreessen made a specific argument recently that cuts to the heart of what's missing in the current AI stack. His point: for an AI agent to move beyond being a chatbot and actually get things done — booking a trip, hiring a specialist, purchasing server capacity, paying an API — it needs to be able to transact independently. Without that capability, agents are, in his framing, trapped in a box. They can give advice. They can't take action.

Watch
Marc Andreessen on AI agents, crypto, and economic identity
Andreessen's case for why agents need permissionless financial rails — and why human infrastructure can't serve them. The argument that this post responds to.
Watch on YouTube →

The reason is structural. Traditional financial systems were built around a single assumption: the account holder is a human being with a government-issued ID, a tax identification number, and a physical address. KYC rules require it. Bank account ownership requires it. Every onboarding flow in every regulated financial institution on earth was designed for a human principal.

AI agents don't have any of that. So they can't hold money. They can't spend money. They can't settle contracts. They are economically inert — which means every task that requires a financial output has to be handed back to a human.

That's the box.

The missing layer

Think about what a genuinely useful AI agent looks like. Not a chatbot that drafts emails — an agent that operates autonomously in the world. It needs to execute on your behalf:

Every one of those actions is a financial transaction. And every one of those transactions currently requires a human to authorise it — not because the agent isn't capable of making the decision, but because the payment infrastructure has no concept of a non-human principal. The agent reaches the point of action and hits a wall.

This isn't a model problem. The models are capable of executing these decisions. It's an infrastructure problem.

Why permissionless rails are part of the answer

Andreessen's argument about crypto infrastructure isn't ideological — it's architectural. Blockchain-based payment rails are permissionless. There's no KYC requirement that filters out non-human entities. There's no account application that asks for a passport number. An AI agent can hold a wallet address and transact natively, without needing a human intermediary to vouch for its identity or co-sign the transaction.

That's the right starting point. But a wallet address alone doesn't close the gap. What agents actually need is the full financial stack built natively for autonomous operation:

Financial identity
A wallet tied to the agent — not borrowed credentials from a human account. The agent is the principal, not a delegate.
Programmable policies
Spending rules enforced automatically — counterparty allow-lists, value caps, time restrictions — without requiring manual sign-off.
Agent-to-agent settlement
When two agents coordinate on a task, they settle directly. No human payment system in the middle, no latency from human approval cycles.
Audit trails
Every transaction logged against the agent identity, the policy it operated under, and the task it was executing — legible to compliance teams without a human signatory.

None of that exists in human financial infrastructure. It has to be built from scratch.

The economic identity of AI

The phrase Andreessen uses — economic identity — is precise. Right now, AI agents have no economic identity. They can think. They can plan. They can execute instructions within software environments. But they are economically invisible — they cannot be a party to a financial transaction, hold assets, or settle liabilities.

This matters more than it might seem. The moment you add financial autonomy to an agent, its capability profile changes completely. It stops being a tool that produces outputs for a human to act on, and becomes a participant in its own right — one that can close the loop on a task without the human re-entering the loop at the point of payment.

That's not a marginal improvement in productivity. It's a different category of system.

Giving agents economic identity is the missing link between agents that advise and agents that act. It's what transforms the AI layer from a productivity tool into a genuine participant in the economy.

Further reading: Autopilots Don't Just Process. They Pay. · Building a Payment Policy Engine for Your AI Agent · Why Non-Custodial Architecture Wins for Agent Finance

Financial infrastructure for AI agents

Wallets, payment policies, programmable spending, agent-to-agent settlement, and treasury — built for autonomous systems. Now onboarding Founders tier partners.

Get early access →