Summary of a new DeepMind paper

Super intriguing idea in this new @GoogleDeepMind  paper – shows how to handle the rise of AI agents acting as independent players in the economy.

It says that if left unchecked, these agents will create their own economy that connects directly to the human one, which could bring both benefits and risks.

The authors suggest building a “sandbox economy,” which is a controlled space where agents can trade and coordinate without causing harm to the broader human economy.

A big focus is on permeability, which means how open or closed this sandbox is to the outside world. A fully open system risks crashes and instability spilling into the human economy, while a fully closed system may be safer but less useful.

They propose using auctions where agents fairly bid for resources like data, compute, or tools. Giving all agents equal starting budgets could help balance power and prevent unfair advantages.

For larger goals, they suggest mission economies, where many agents coordinate toward one shared outcome, such as solving a scientific or social problem.

The risks they flag include very fast agent negotiations that humans cannot keep up with, scams or prompt attacks against agents, and powerful groups dominating resources.

To reduce these risks, they call for identity and reputation systems using tools like digital credentials, proof of personhood, zero-knowledge proofs, and real-time audit trails.

The core message is that we should design the rules for these agent markets now, so they grow in a safe and fair way instead of by accident.

That is from Rohan Paul, though the paper is by Nenad Tomasev, et.al.  It would be a shame if economists neglected what is perhaps the most important (and interesting) mechanism design problem facing us.