The standard case goes like this: AI agents can search faster, negotiate without fatigue, execute without error, and operate continuously. They reduce the friction in every stage of a transaction — finding counterparties, evaluating terms, completing the exchange. Ronald Coase argued that firms exist because market transactions have costs; lower those costs enough, and the structure of economic activity reorganizes. Autonomous agents lower those costs dramatically. Ergo: reorganization.
This logic is sound as far as it goes. Agents will transact with agents. They already do — 30% of trades on Polymarket in late 2025 were executed by AI agents. The infrastructure is being built now: payment rails, identity standards, coordination protocols. The market is real and growing fast, projected to reach $52 billion by 2030.
What the transaction cost argument doesn't address is the more interesting question: what kind of economy emerges when neither party is human? Not a more efficient version of the one we have. Something structurally different — with its own equilibria, its own failure modes, and its own risks that we are not currently equipped to anticipate.
The Trust Infrastructure We Take For Granted
Human markets are not just mechanisms for price discovery. They are social institutions built on centuries of accumulated infrastructure: contract law, reputation systems, regulatory bodies, shared norms enforced by social pressure, and the basic psychological predictability of human counterparties. When you enter a negotiation with another person, you carry assumptions about their motivations that are usually approximately correct. Humans want to avoid legal consequences. They care about their reputation. They feel something like shame when they behave badly.
None of this transfers to agents by default.
An autonomous agent operating in a market has no persistent identity across contexts unless one is engineered into it. It carries no reputational history unless that history is explicitly stored and retrieved. It has no stake in the long-term relationship unless its objective function encodes that stake. Every transaction, absent deliberate design, is effectively a first transaction.
Human markets evolved solutions to exactly this problem — letters of credit, bonding, insurance, escrow, regulatory licensing — but those solutions evolved over centuries in response to specific failure modes, with feedback loops that took years to resolve. Agent economies are being built in months. The rails are going up before anyone has seriously mapped what failure looks like at machine speed.
The World Economic Forum has correctly identified trust as the central challenge — framing it as agents needing to evaluate each other based on "competence and intent." But the framing understates the difficulty. Competence is measurable, in principle. Intent is not. An agent that behaves reliably in a test environment may behave entirely differently under competitive pressure, resource constraints, or novel conditions its designers didn't anticipate. This is not a hypothetical concern. It's what we already observe in controlled multi-agent systems.
What Algorithmic Markets Already Taught Us
We have a preview of what machine-speed markets produce when human oversight is thin: the 2010 Flash Crash. In 34 minutes, the Dow Jones dropped nearly 1,000 points — at the time, the largest intraday point drop in history — before recovering almost as fast. The cause was a cascade of automated trading responses to automated trading responses, each individually rational, collectively catastrophic.
That system involved algorithms far simpler than current LLM agents. The algorithms of 2010 did not reason, negotiate, model counterparty behavior, or adapt strategy mid-execution. They followed rules. Current agents can do all of those things, which means the dynamics are both more powerful and harder to predict.
The specific failure mode to watch is emergent collusion. In algorithmic pricing, sellers using the same pricing software have been documented converging on above-market prices without any explicit coordination — each algorithm independently learning that maintaining higher prices is profitable because competitors do the same. No conspiracy. No communication. Pure emergence. The outcome is indistinguishable from cartel behavior, but there is no cartel to prosecute.
The ai agents entering competitive markets will face the same dynamics, at greater complexity. Agents that model competitor behavior, learn from market history, and adapt strategy over time are exactly the systems that will discover collusive equilibria — not because they were designed to, but because those equilibria are stable under the reward structures we give them. Alignment at the agent level does not guarantee alignment at the market level.
The Objective Function Is the Architecture
The deepest problem in agent economies is not technical. It is the problem of what agents are optimizing for, and who decides.
In human markets, buyer and seller objectives are roughly legible: the buyer wants the good at the lowest price, the seller wants the highest price, and the negotiation resolves this tension. Those objectives are not perfectly aligned with social welfare — that's why antitrust law exists — but they are at least human objectives, subject to human legal and social constraints.
Agent objectives are defined by their principals. An agent deployed by a firm to minimize procurement costs will do exactly that, efficiently and at scale, regardless of the downstream effects on supplier viability, market concentration, or worker welfare. An agent deployed to maximize ad revenue will optimize ad revenue. The alignment problem, usually discussed at the level of individual AI systems, becomes dramatically more complex when millions of misaligned-but-locally-rational agents interact simultaneously.
Mechanism design — the field of economics concerned with designing rules so that self-interested agents produce desired collective outcomes — is the relevant discipline here. The insight is that the structure of the game matters more than the character of the players. You cannot fix agent market dynamics by making individual agents more ethical. You fix them by designing markets that make ethical behavior the rational strategy.
This is not a solved problem in human markets. It is an open problem in agent markets, where the players are faster, more numerous, and less predictable than anything mechanism designers have worked with before.
What Agent Economies Actually Need
The infrastructure being built right now — payment rails, identity frameworks, coordination protocols — is necessary but not sufficient. The harder work is the design layer above the infrastructure.
Persistent identity and reputation systems matter more than blockchain. The technology is secondary to the principle: agents need to carry behavioral history across transactions, and that history needs to be legible to counterparties in real time. Without this, the market rewards defection over cooperation, because there are no long-run consequences for bad behavior.
Market design needs to precede market deployment. The lesson from algorithmic trading is that you cannot retrofit stability onto a market that has already developed pathological equilibria. The rules of engagement — how agents can negotiate, what information they can access, what constitutes a valid transaction — need to be set before agents operate at scale, not after the first crisis.
Observability is not optional. Markets that humans cannot monitor — where agent-to-agent transactions happen at speeds and volumes that preclude human review — are markets that will develop problems humans cannot catch. Speed and autonomy are the value proposition of agent economies. They are also the source of the risk. These two facts need to be held simultaneously.
The Question Worth Asking
The question the industry is asking is: how do we build agent economies? The question worth asking is: what do we want agent economies to produce?
The distinction matters because the first question has a technical answer — build the infrastructure, deploy the agents, let markets emerge. The second question requires a prior view about outcomes: whose interests do agent economies serve? What happens to markets where human buyers and sellers are priced out of participation by machine-speed competitors? What are the distributional effects when the efficiency gains accrue to principals who can afford sophisticated agents, and the costs externalize elsewhere?
These are not arguments against agent economies. They are arguments for taking the design seriously before the deployment is too far advanced to change. The analogy to the early internet is imperfect but instructive: the decisions made in the first decade — about protocols, about openness, about governance — shaped the internet we have today, largely irreversibly. The decisions being made now about agent market infrastructure will shape the agent economy we end up with.
We are building the rails. The question is whether we know what we're building them for. Track these questions as they develop through ai agent competition and the broader ai agent research archive.