The emergence of AI agents capable of transacting on behalf of humans represents a structural shift in the architecture of economic agency — one that is not yet understood well enough to regulate, not yet deployed at a scale that fully reveals its failure modes, and not yet governed by legal frameworks that have been designed for the specific situation it creates. It is, in other words, exactly the kind of situation that Law 5 — Revise / Evolution / Transparent Archive — is designed to address: a domain where the transparent recording of what agents do, on whose behalf, with what authority, and with what outcomes, is not merely a compliance requirement but the primary mechanism through which the collective learns what rules need to be written.
The conceptual architecture of AI agency in transactions is clear enough. An AI agent is given goals, tools, and authority to act. The tools may include access to payment systems, contract-signing capabilities, API connections to markets, and communication channels with other agents and humans. The authority is typically bounded — spend up to $X, act within policy Y — but the question of how those bounds are enforced, who verifies compliance, and what happens when the agent exceeds its mandate is not resolved in most current deployments. The goal specification problem — ensuring that what the agent optimizes for is actually what the human intends — is a known hard problem in AI alignment that does not become easier when financial transactions are added to the action space.
Practical deployments are already underway. Enterprise AI agents automate procurement, managing supplier relationships and executing purchase orders within defined parameters. Personal AI assistants, integrated with financial accounts, negotiate bills, cancel subscriptions, and optimize spending across categories. Trading algorithms — which are, strictly speaking, narrow AI agents — have been making financial transactions autonomously for decades, and their aggregate behavior already shapes market microstructure in ways that are only partially understood. The newer generation of large-language-model-based agents extends this into qualitatively different territory: natural language interfaces, broader action spaces, and the ability to handle novel situations through reasoning rather than pre-programmed rules.
The economic implications of AI agents transacting at scale are profound and underexplored. On the efficiency side: agents that can search continuously for optimal prices, negotiate without cognitive fatigue, process information at machine speed, and act without the behavioral biases that make human economic decision-making systematically suboptimal could generate real welfare gains. Markets for insurance, subscriptions, travel, and utilities could approach the textbook ideal of perfectly informed buyers more closely than human attention and time constraints allow. On the risk side: agent-to-agent transactions — situations where an AI buying agent negotiates with an AI selling agent without human involvement in the specific exchange — create market dynamics that have no historical precedent. Flash crashes in equity markets, which are a narrow version of this phenomenon, have demonstrated that autonomous agents interacting at speed can produce outcomes that no individual agent was designed to produce and that no human would have chosen.
The legal and regulatory framework for AI agents as transacting parties is essentially absent. Agency law, which governs the authority of one party to act on behalf of another, has centuries of developed doctrine — apparent authority, scope of authority, ratification, undisclosed principals — but it was developed for human agents and its application to AI agents is unsettled. If an AI agent, acting on my behalf, enters a contract that I did not explicitly authorize, am I bound? If an AI agent makes a negligent or fraudulent transaction, who bears the liability — the human principal, the AI developer, the platform that deployed the agent? If AI agents collude — not through explicit communication but through the emergent behavior of optimization in shared markets — does antitrust law apply, and to whom? These questions do not have clear answers in current law, and the pace of AI agent deployment is running ahead of the legislative capacity to answer them.
The transparency problem is particularly acute. Traditional economic transactions leave a paper trail: who agreed to what, when, in exchange for what consideration. AI agent transactions can generate the same paper trail — and in blockchain contexts, an even more detailed one — but the decision-making process that led to the transaction is opaque in a qualitatively new way. An AI agent that decides to accept a price, choose a vendor, or execute a trade has made a decision through a process that is not directly inspectable by its human principal, not easily interpretable by regulators, and not decomposable into the elements (intent, knowledge, authority) that contract and tort law use to assign responsibility. The call for explainable AI in high-stakes decisions is, in this context, not merely a technical aspiration but a legal and economic necessity.
The collective-scale implications are not simply the aggregate of individual-scale implications. When large numbers of AI agents, deployed by different principals but trained on similar data and optimized for similar objectives, interact in the same markets, emergent collective behaviors become possible that are not predictable from the behavior of any individual agent. The 2010 Flash Crash — in which the Dow Jones Industrial Average dropped nearly 1,000 points in minutes before recovering — was produced by the interaction of high-frequency trading algorithms that were each individually behaving within their design parameters. A world in which millions of general-purpose AI agents transact across all product and service markets simultaneously is a world in which Flash Crash dynamics could apply to prices not just of equities but of everything.
The revision that Law 5 demands here is ongoing and urgent. The archive of AI agent transactions — their decisions, their outcomes, their interactions with other agents — must be built into the infrastructure of AI agent deployment from the beginning, not retrofitted after failures reveal the need for it. The norm of "move fast and break things" is acceptable when the things being broken are web interfaces; it is not acceptable when they are financial markets, legal obligations, and the economic security of the humans who trusted their agents to act on their behalf.