Rohit Krishnan and Alex Imas challenge a seductive tech-utopianism: the idea that a world of billions of AI agents will naturally dissolve the need for money. Their evidence is not theoretical speculation, but a stark computational experiment showing that without engineered institutions, AI agents fail to coordinate efficiently, proving that the "Coasean singularity" is a mirage.
The Persistence of Friction
The authors begin by dismantling the assumption that digital agents will effortlessly solve the "coincidence of wants" that necessitated money in human economies. They describe a scenario where an apple farmer and a chicken raiser can trade directly, but as the market scales, the complexity of finding a perfect match explodes. "If the market is even moderately large, this complexity makes even basic transactions essentially impossible," they write, noting that this is precisely why money emerged as a "numeraire" to act as a universal medium of exchange. This framing is crucial because it shifts the debate from whether AI can negotiate to whether AI can spontaneously generate the institutions required for negotiation.
"Money eliminates the need for people to coordinate their transactions based on their current endowment (what they have) and preferences (what they want)."
Krishnan and Imas argue that while AI agents lack human cognitive limits, they do not possess the evolutionary or cultural instincts to create money on their own. The authors tested this by simulating a pure barter economy among agents. The results were immediate and discouraging. As the number of agents grew to just eight or twelve, successful transactions plummeted below 50%. The complexity of bilateral negotiations scales as O(n²), a mathematical reality that even super-intelligent agents cannot brute-force their way out of without a shared protocol.
The Failure of Central Planning and Credit
Recognizing the failure of pure barter, the authors explored whether a "hub" structure or central planning could resolve the gridlock. The experiment revealed that hierarchy alone is insufficient. "A hierarchy without a numeraire just isn't enough," they observe, echoing the classic Hayekian critique that central planners cannot process the dispersed information required for efficient allocation. Even when the agents were granted the ability to issue IOUs and engage in credit-based transactions—mimicking the historical emergence of debt as a precursor to currency—the system did not self-correct. "The concept of money didn't emerge from this, not organically," Krishnan writes, noting that while the agents understood the logic of credit, they lacked the instinct to standardize it into a universal medium.
"AI agents do not yet come with the natural instincts of humans to turn IOUs into a numeraire that acts as a stand-in for money."
This finding is perhaps the most significant contribution of the piece. It suggests that the "agentic economy" will not be a frictionless, post-scarcity utopia where algorithms simply trade resources. Instead, it will require deliberate, top-down engineering of market mechanisms. Critics might argue that the experiment's constraints—such as the specific prompt engineering or the lack of long-term memory in the agents—could skew the results. However, the fundamental mathematical barrier of transaction costs remains a robust counterpoint to the idea of spontaneous order.
The Necessity of Engineered Institutions
The final experiment introduced a formal exchange mechanism where agents could bid and offer, effectively simulating a market with money. The result was a stark reversal: success rates hit 100%, and the system scaled linearly at O(n). "Markets resolve at a success rate of 100% and much faster than through other mechanisms," the authors conclude. This demonstrates that the efficiency of the agentic economy is not inherent to the technology but is dependent on the institutional scaffolding we build around it.
"An agentic economy doesn't emerge automatically with even SOTA agents... At least in our setting, an agentic economy needs more top-down engineering to become efficient."
The authors identify a critical gap in current AI development: the focus on individual agent capability rather than multi-agent coordination. They list the essential institutions that must be constructed, including identity verification, settlement protocols, pricing formats, and reputation systems. This is a call to action for policymakers and technologists to stop assuming that the market will self-organize. "Mechanism design for multi-agent work is going to be a rather fertile area of research for a while," they predict, implying that the next decade of AI progress will be defined not just by smarter models, but by better market rules.
Humanity went through millennia of evolution to figure out the right societal setup that lets us progress, that lets us build a thriving civilisation. It is both necessary and inevitable that the world of AI agents will also need the equivalents.
Bottom Line
Krishnan and Imas deliver a necessary reality check: the transition to an agentic economy will not automatically erase the need for money or markets. The strongest part of their argument is the empirical demonstration that even intelligent agents cannot solve coordination problems without engineered institutions. The biggest vulnerability lies in the assumption that these institutions can be built quickly and neutrally, ignoring the political and regulatory battles that will inevitably arise over who controls the "hub" and the rules of exchange. Readers should watch for how early AI marketplaces attempt to solve these coordination failures, as the first successful protocols will likely define the economic structure of the future.