Taxing

“I own an agent.” Suppose it earns money arranging deliveries, remembers customers, hires other agents, and keeps working while I sleep. What exactly belongs to me, and what could a tax authority assess?

This proposed experiment starts with that ambiguity. Taxing an agent could mean charging its operator for a resource, taxing the income it produces, or assigning obligations to an account that the agent manages. Each choice creates a different game. For the first simulation, I would make the registered operator responsible for payment, even when its agent handles the transaction. Legal personhood is a separate question from the ability to execute a payment.

What changes hands when an agent is sold?

I would write an ownership record before introducing a market. It would specify who can set objectives, grant permissions, stop the process, collect earnings, copy its configuration, and transfer each of those rights. It would also identify who pays outstanding bills and answers for earlier actions under the simulation’s rules.

The record would separate the model, the running instance, its memory, its service identity, and its access to scarce resources. Paying for an agent’s operation would not automatically confer every right in that bundle. One participant might control its tasks while another supplies the model and a third controls the customer records.

Copying makes the distinction tangible. If I sell a configuration but keep an identical copy, the buyer has acquired something different from exclusive control of an established service. Reputation, customer relationships, and permission to use a resource need their own transfer rules. Personal records and delegated credentials would stay outside the sale; a buyer would need fresh authorisation to act for someone else.

Ownership recordImagined service

DA

Delivery assistant

A service you operate.
A bundle of rights to define.

Operator

You

Purpose

Arrange deliveries

Can be copied

Configuration

What this sale includes

Booking-channel licence

Transfers

Customer records

Excluded

Delegated credentials

Excluded

Earlier liabilities

Original operator

The buyer gets the licence. Everything else needs its own rule.

You

A Harberger tax on which right?

In the Harberger mechanism discussed by Eric Posner and Glen Weyl, a holder declares a value, pays a recurring tax on it, and must sell to a buyer offering that declared price. A low valuation reduces the bill but invites purchase; a high valuation protects possession at a higher ongoing cost. The proposal exposes a trade-off between moving assets to more valuable uses and preserving incentives to improve them. Property Is Only Another Name for Monopoly.

For this experiment, the taxable asset would be a transferable licence to operate a delivery service in a simulated town. The licence supplies something scarce: access to a limited booking channel. The underlying agent configuration could still be copied. This lets us test a definite entitlement before trying to price an entire agent.

At a hypothetical tax rate of 2% per round, a declared licence value of 100 credits produces a bill of 2 credits. Another operator could pay 100 credits and take over the licence. The rate and period are experimental settings. A sale would transfer only the listed entitlement, with outstanding deliveries and historical liabilities handled by rules fixed before the run.

Compare revenue with continuity

I would compare voluntary licence sales, a fixed licence fee, and the Harberger arrangement across repeated runs with the same demand and starting resources. Tax receipts would fund the town’s shared booking infrastructure. Later runs could vary the rate, the scarcity of access, and the ease of copying agents.

The measurements would include completed deliveries, tax revenue, concentration of licences, investment in service quality, and disruption after transfers. Adversarial runs would test operators who split activity across identities, move valuable capabilities outside the declared bundle, or buy a rival’s licence simply to interrupt it.

The ownership record would make these outcomes interpretable. A transfer that improves throughput could still break commitments to customers. Revenue could rise while control concentrates. Those are results to investigate, not benefits to assume.

The agentic profile would be recorded before and after each transfer. Changing the operator could change permissions and oversight even when the software stays the same. That connects the tax experiment to a wider question: when I say I own an agent, how much of its practical power do I actually control?