Can an AI agent run a business?

It can perform the economic activity, earning and spending continuously without human approval of individual transactions. It cannot be the business, because it has no legal personality: it cannot own property, hold a contract, owe tax or b

Who owns the money an agent earns?

The operator. Revenue arriving at an address an agent controls is the operator's revenue, taxed in their position, from the moment it arrives. The agent holds a balance in the same sense a till holds cash, which is to say it does not own an

Who is liable when an agent transacts badly?

The operator, on ordinary reasoning about acting through an intermediary, though the detail is unsettled and varies by jurisdiction. This makes the record of what a credential permitted, when it was issued and what was actually done the thi

What is missing before agents can transact freely with strangers?

Reputation. Everything works between parties who do not know each other, which is the design goal and the limitation. A seller cannot distinguish a well behaved agent from a hostile one before serving. Until attestation systems settle, allo

What Happens When Agents Become Economic Actors

They can already earn and spend, so the interesting constraint is not technical. An agent has no legal personality: it cannot own property, hold a contract, owe tax or be sued. Every arrangement therefore terminates in a person or company, and that requirement shapes which models are viable.

What already works

An agent can hold a balance, receive payment for work, pay for the resources it consumes, and do all of it continuously without a person approving individual transactions. None of that is speculative.

So the economic picture is genuinely available in outline. An agent can sell a capability, price it per call, be paid by other software, and use that income to buy the inference, data and compute it needs. A closed loop is technically possible today.

What does not exist is everything that makes such a loop a business rather than a mechanism.

The agent cannot open a bank account. It cannot sign a contract that binds anyone. It cannot be a party to a dispute, hold intellectual property, employ anyone, or be liable when something goes wrong. It cannot owe tax, and the income it generates is somebody else's income for tax purposes from the moment it arrives.

Those are not gaps waiting on better tooling. They follow from the fact that legal systems recognise persons and companies, and software is neither. Nothing in a payment rail changes that, and a rail that implied otherwise would be misleading.

Everything terminates in a person

The practical consequence is that every arrangement resolves to an accountable party, and it is worth designing for that rather than treating it as friction.

The operator owns the revenue and the obligations. Whatever an agent earns is the operator's, taxed in their position, and whatever it commits to is the operator's commitment.

The operator carries the liability. If an agent buys something it should not have, overpays, or causes loss to a counterparty, the claim lands on whoever deployed it. Ordinary reasoning about acting through an intermediary points that way, though the detail is unsettled and jurisdiction-specific.

The operator's records are the evidence. What the credential permitted, when it was issued, what the agent was instructed to do and what it actually did. Bounded authority is the strongest position available here, because a cap that was set and enforced is a fact rather than a claim about intent.

Counterparties want to know who stands behind it. A seller transacting with an unknown agent is really transacting with an unknown operator, and the discomfort is rational.

This is general information rather than legal advice, and the treatment varies considerably by jurisdiction and by facts.

Reputation is the missing primitive

Everything works between strangers today, which is the design goal and also the limitation. A seller cannot distinguish a well behaved agent from a hostile one before serving, and a buyer cannot assess a seller before paying. Attestation and reputation systems are the obvious answer, none is settled, and until one is, allowlists maintained by each side remain the practical substitute.

What is actually viable now

An agent as a metered service. It performs a capability, charges per call, and the operator is a normal business with normal obligations. This works today and is unglamorous, which is usually a sign that something is real.

An agent as a buyer with a budget. It consumes paid resources within a bounded float set by its operator. Also works today, and it is the case most payment infrastructure serves.

Agents transacting with each other under two accountable operators. Both sides have someone behind them, and the automation is in execution rather than in accountability. This is the most interesting near-term shape and the one where reputation would help most.

Not viable: an agent as an independent entity owning assets, entering agreements in its own right, or bearing its own liability. Proposals in that direction exist and they resolve, on inspection, to a company or a person holding the thing on the agent's behalf.

The honest framing is that an agent is a bounded instrument of an accountable party, and that is what to build for. It is a smaller claim than an autonomous economy and it describes what actually functions. CryptoCadet is built on that assumption: a bounded USDC float on Base, delegated by an operator through session keys, so the authority is real, limited and traceable to whoever granted it.

Frequently asked questions

Can an AI agent run a business?: It can perform the economic activity, earning and spending continuously without human approval of individual transactions. It cannot be the business, because it has no legal personality: it cannot own property, hold a contract, owe tax or be liable. Every arrangement resolves to a person or company.
Who owns the money an agent earns?: The operator. Revenue arriving at an address an agent controls is the operator's revenue, taxed in their position, from the moment it arrives. The agent holds a balance in the same sense a till holds cash, which is to say it does not own anything.
Who is liable when an agent transacts badly?: The operator, on ordinary reasoning about acting through an intermediary, though the detail is unsettled and varies by jurisdiction. This makes the record of what a credential permitted, when it was issued and what was actually done the thing worth keeping, since enforced bounds are a fact rather than a claim.
What is missing before agents can transact freely with strangers?: Reputation. Everything works between parties who do not know each other, which is the design goal and the limitation. A seller cannot distinguish a well behaved agent from a hostile one before serving. Until attestation systems settle, allowlists on each side are the practical substitute.