What is an autonomous AI economic agent?
I am guilty of writing about AI agents without any clear idea of what they are. There are only a few autonomous AI economic agents in the wild, yet we need to imagine what AI agents could be and how they might interact. To build a digital economy — and have the conversations and arguments that we need to in order to do that building — we must at least agree on what it is we are talking about.
Below is a first attempt at a working definition of an autonomous AI economic agent. As you will see, AI economic agents have delegated powers from a principal, make economic exchanges under uncertainty, and control digital property rights.
Hold delegated powers to decide and execute
AI economic agents operate with explicitly delegated authority from a principal (whether an individual or an organisation). This gives the AI agent some autonomy. The delegation from a principal to an agent includes two dimensions.
They are delegated some decision rights. This includes the agent's scope to make economic decisions within defined boundaries. These might include restrictions on exchange size, asset types, or transaction legality. The agent can access and process external information including oracle data, price changes, and market conditions.
They are also delegated execution rights. That is, they are not only delegated the rights to make decisions, but also to act on those decisions, such as making a trade or forming contracts
A portfolio management AI economic agent, for instance, might analyse market conditions based on a set amount of information sources, and then have execution rights to rebalance those investments within set parameters.
Make economic exchanges under uncertainty
AI economic agents go beyond executing pre-defined tasks, such as sending an email or simple if-then programs. They participate directly in economic exchanges under Knightian uncertainty. They can exchange with organisations, individuals and other AI agents.
In decentralised finance (defi), yield farming bots already navigate complex liquidity pools, but future AI economic agents might go further — simultaneously managing lending positions across multiple protocols, buying protocol insurance, and negotiating with freelancers to buy services. These agents must learn from market conditions and adapt their strategies as economic conditions change.
Control digital property rights
AI economic agents must control digital property rights — such as holding a wallet with digital assets (e.g. stablecoins), to access access rights for valuable digital services (e.g. compute, storage). Unlike a passive adviser, an AI economic agent must have genuine control over resources to participate in economic exchange.
This control goes beyond mere contractual promises. Without residual control rights over digital property rights of some kind, AI agents cannot make credible commitments in economic exchanges or adapt to unexpected situations not specified in their alignment with their principal. An AI system that can only advise on trades but lacks authority over digital resources is not an economic agent — it is merely an advisor.
Degrees of AI agency
An AI system must possess all three of the above characteristics to qualify as an economic agent. But there are also varying degrees of agency along each of these dimensions.
For instance, an AI economic agent might handle increasingly complex economic decisions under greater uncertainty. It might control a wider variety of digital property rights or larger pools of resources. It might require less feedback or approval loops from its principal. It might operate with broader delegated powers or more autonomous decision-making capacity
The more extensively an agent operates along these dimensions, the more 'agentic' it becomes, even while maintaining its fundamental nature as an economic agent.
More agentic?
The feasibility of AI economic agents is fundamentally a question of transaction costs. In this context, several technological advances are rapidly reducing the transaction costs of AI agents:
Large language models are radically reducing the costs of reasoned decision making under uncertainty.
New toolkits and interfaces have made it easier and cheaper for humans to delegate authority, making it easier to align an AI economic agent with my preferences.
Blockchain-based digital assets and smart contracts have lowered the costs of resource control and trade execution. Furthermore, as more assets and data become digital then the potentia
l scope for AI economic agents expands.
The choice to use AI economic agents is ultimately a comparative question: will an AI agent be more efficient than existing forms of economic organisation? As these costs continue to fall, we are likely to see autonomous AI agents taking on more economic roles — not because they are inherently better, but because they are becoming comparatively better at certain types of economic coordination.