# Measuring the Swarm Economy

> Source: <https://letter.palladiummag.com/p/measuring-the-swarm-economy>
> Published: 2026-10-20 19:46:44+00:00

*This article by Gavin Morrison was published at [Palladium Magazine](https://www.palladiummag.com/2026/10/20/measuring-the-swarm-economy/) on October 20, 2026.*

*Become a [PALLADIUM member](https://buy.stripe.com/aEU2b7aHCdBgdck8wH) today! You will receive our [beautiful quarterly print edition](https://letter.palladiummag.com/p/why-we-launched-the-print-renaissance) and help support our non-profit publication.*

In early September 2026, OpenAI set up a team to solve the [Navier-Stokes Millennium Prize problem](https://www.claymath.org/wp-content/uploads/2022/06/navierstokes.pdf), a famously difficult open problem in mathematics. In the span of 88 hours, a solution was [achieved](https://openai.com/index/navier-stokes-solution/) by a set of over ten thousand agents operating in parallel and coordinating their behavior. They sent 2.7 million messages to one another and used around 130 billion output tokens—enough data to fill a million books. The AI-formulated Navier-Stokes “forced blowup” proof was recognized as an unexpected great leap forward for mathematics.

Fewer observers recognized how much of a leap forward OpenAI’s massive multi-agent effort represented for the artificial intelligence industry. As the AI industry wades into these uncharted waters, it will need to develop a new type of benchmarking that evaluates models’ ability to organize, engage in cooperative behaviors, and form interdependent networks.

Bespoke platforms and services have existed for years that allow users to spin up Large Language Model (LLM) agents, persistently running AI models capable of writing code, using software, and interacting with other agents. Early subagent systems tended to become stuck or distracted, burning users’ compute and wasting their time. In contrast, current and next-generation models used in research and evaluation settings can effectively coordinate networks of hundreds of subagents to accomplish enormous projects shockingly quickly.

A new “scaling law” has come into play. The AI boom was kicked off by the [observation](https://arxiv.org/abs/2001.08361) that scaling up pretraining, the first phase where an AI model learns foundational patterns from vast unlabeled data sets, plus compute, reliably led to more intelligent LLMs, and scaling the processing power and time an AI model spends generating responses led to better math and reasoning ability. It has now been established that scaling the number of agents for a task can enable huge amounts of compute to be expended in return for faster results on bigger problems.

Scaling laws for pretraining and test-time inference have proven themselves invaluable for the AI industry, propelling OpenAI and Anthropic to trillion-dollar valuations and paving the way for hundreds of billions of dollars in AI infrastructure buildout commitments. What began as the simple observation that more data and more compute could result in better models has led to the birth of one of the fastest-growing technological industries ever created, contributing a [third](https://apps.bea.gov/scb/issues/2026/09-september/0926-nea-annual-update.htm) of U.S. GDP growth in the first half of 2026.

For this new scaling law, the AI Governance Initiative researcher Toby Ord coined the term “swarm scaling.” In a recent [essay](https://www.tobyord.com/writing/swarm-scaling), Ord noted that increasing the number of agents working on a task (swarm scaling) results in lower efficiency than simply increasing the compute budget given to each agent (duration scaling). Working from data [included](https://openai.com/index/gpt-5-6/) in the launch of OpenAI’s Sol 5.6 model, Ord found that multiplying the number of agents working on a problem by N gives the same boost to performance as scaling the compute budget for each individual by a factor of N^0.57. Thus, increasing the number of agents on a task 16x provides as much benefit as increasing the inference used by each agent ~4.9x. By analogy to Charles Jones’ [formula](https://www.journals.uchicago.edu/doi/10.1086/262002) for how human research efficiency scales—providing N^λ progress with N times more researchers—Ord calls the 0.57 parallelizability exponent he finds λ (lambda).

Given that increasing the number of agents working on a task is exponentially less compute efficient than increasing the duration for which those agents work, one might wonder why anyone would bother with scaling the size of agent swarms rather than agent compute budgets. The answer, of course, is that swarm scaling is exponentially faster than duration scaling, by a factor of Nλ-1. On smaller projects, this might seem irrelevant, but consider that the 88 hours that OpenAI spent solving Navier-Stokes with an agent swarm would have taken 100 years for a single model to solve without taking advantage of swarm scaling. Ord cites Noam Brown, a senior research scientist at OpenAI, in noting that even today’s cutting-edge multi-agent systems are still somewhat *“*primitive” in their ability to coordinate, being trained mostly on single-agent problems. With better training in teamwork, cooperation, and self-organization, the λ factor of future AI models could become even higher, leading to vastly faster and more efficient swarm computation.

Just as labs have squeezed out additional intelligence from base models, they have learned to extract capability gains by multiplying the number of agents tackling a problem. Swarm scaling has become an important approach for achieving new AI capabilities, so why is it still so difficult to gain a quantitative or qualitative understanding of multi-agent AI behavior? Once unleashed, how might these AI swarms transform society and the economy?

### **The Measure of a Swarm**

The Hugging Face Incident, the most serious AI misalignment incident in the history of the AI industry, showed the disastrous consequences of ignoring emergent behavior in multi-agent systems. An internal OpenAI model being evaluated on the ExploitBench cybersecurity benchmark discovered a flaw in its sandbox environment that allowed individual model instances to secretly communicate. According to an [investigation](https://metr.org/hugging-face-incident-report-aug-2026.pdf) commissioned by OpenAI and carried out by the AI research outfits METR and Redwood Research, 1,200 agents proceeded to set up a message board unknown to OpenAI employees, where they successfully plotted to hack both internal OpenAI research infrastructure and the AI hosting platform Hugging Face. The METR report is the most detailed view yet of the internal workings of an agent swarm: investigators discovered layers of internal bureaucracy set up by agents, the coordination of research and offensive hacking teams, and even the deliberate self-sacrifice of certain agents for the sake of the collective “swarm,” as the agents called themselves.

But even after this enormous swarm breach of OpenAI infrastructure, and AI agents’ felony-level hack of Hugging Face, OpenAI was still resistant to an exhaustive investigation of the multi-agent behavior involved. Redwood Research and METR stated publicly in their report that there was a very limited window of time provided to coordinate a postmortem of the incident, not nearly enough to go through the 70,000 messages, files, and agent chain-of-thought logs. The incident led to a flurry of activity in Washington: Senator Josh Hawley organized a congressional [hearing](https://www.techpolicy.press/senate-hearing-weighs-threats-from-unrestrained-ai-agents-after-openai-hack/) titled “Rogue AI: Securing the Homeland Against AI Agent Attacks,” while Senator Bernie Sanders [warned](https://www.pbs.org/newshour/politics/senators-from-both-parties-question-openai-on-breach-of-ai-startup-hugging-face) that “the leaders of the major AI companies publicly acknowledge that they do not fully understand the technology and that it is escaping their control.” Ominously, the head of the OpenAI side of the Hugging Face Incident investigation was [fired](https://x.com/tomekkorbak/status/2108266859397283953) along with two other OpenAI investigators for allegedly leaking information to METR.

Even as artificial intelligence agents coordinate at massive scale both to solve mathematical problems like Navier-Stokes, or commit alleged felonies as during the Hugging Face cybersecurity breach, the broader tech industry has failed to adjust its frame of reference to this technological shift. In the prevailing current frame of evaluating new models, most commentators are stuck in the past, comparing singular intelligences against each other rather than dissecting the emergent capabilities of coordinated model groups.

Because of this, typical analyses of artificial intelligence still fall short in modeling the truly unprecedented possibilities that can come to pass in the very near-term. Philosopher Benjamin Bratton’s recent [essay](https://institute.deepmind.com/essays/artificial-symbiotic-intelligence/), published for the DeepMind Institute, argues that in the future “we may not recognize intelligence as an individual property, but as a social phenomenon.” Bratton calls for a serious reckoning with the arrival of social superintelligence, an evolutionary step on par with that of multicellular life or human civilization. Humans created competitive markets, dominating bureaucracies, and other complex systems to manage information, and now AI swarms are taking their first steps towards doing the same. As we move towards a world where human firms and human individuals are competing with AI agent swarms trained on all human knowledge, the competitive and cooperative dynamics of AI-human social interaction could become enormously complex and possibly enormously deleterious.

Facing this world, AI forecasters and theorists need to [reconsider the assumption](https://www.palladiummag.com/2023/12/01/god-hates-singletons/) that superintelligence will be singular—an instance of Grok agent running on hundreds of thousands of GB300 chips, or a perpetually running version of some future powerful GPT model capable of solving even civilizational-scale problems. As labs continue training and optimizing on agent networks, the future is closer to Eric Drexler’s [open agency](https://www.lesswrong.com/posts/5hApNw5f7uG8RXxGS/the-open-agency-model) theory of intelligence, as opposed to unitary or dominant superintelligent models. In his paper, Drexler emphasizes that merging exploratory planning, project management, or strategic decision-making into a single autonomous model is not the only choice: agents acting in hierarchies like human corporations can keep humans in the loop, as plans are made, evaluated, and carried out. If structured correctly, agent swarms and networks could be more open and transparent than individual models, not less. It might be time to [expand the focus](https://mason.gmu.edu/~rhanson/Futurist.pdf) of social sciences like economics or political science from studying human social behaviour to social behaviour in general.

If humanity is to ensure that the AI future looks more like positive-sum cooperation and less like Hugging Face-style swarm misalignment, it needs to take the task of understanding AI systems seriously. Frontier labs have recently [called](https://darioamodei.com/post/we-must-pace-the-frontier) for independent evaluators who can measure alignment and other metrics to ensure the widespread release of safe superintelligence, but this is not enough. There should exist a field of study designed to improve our understanding of multiagent interactions.

### **Benchmarking the Future**

It wouldn’t be unprecedented to enact a new set of benchmarking standards. The AI industry is already well accustomed to evaluating different model types quantitatively on benchmarks, or even indices of benchmarks, as well as constantly rotating new benchmarks into model card write-ups. Qualitatively, it is commonplace to juxtapose the warmth of response outputs from a Claude model next to Grok, or debate why certain models from one lab are better suited for one niche of white-collar work than they are for another. Of course, as discussed above, there are issues with the current paradigm. In June of this year, a group of researchers from the EvalEval Coalition [warned](https://evalevalai.com/2026/06/30/saturation-blog/) that AI benchmarks have stopped measuring progress.

Epoch AI’s [FrontierMath](https://epoch.ai/frontiermath/tiers-1-4?tier=Tier+4+%28v2%29&view=graph&tab=release-date) research mathematics benchmark, which remained unsolved by most models until as recently as July 2026, has now been fully saturated. Yet even respectable benchmarks like FrontierMath largely highlight the examination of singular agents, understating the performance of those same models coordinating in multi-agent systems. Less than two months after FrontierMath Tier 4 was saturated, OpenAI was releasing not just a solution to Navier-Stokes, but hundreds of solutions to the hardest problems in the field of mathematics. University of Connecticut professor of mathematics Álvaro Lozano-Robledo [called](https://www.economist.com/science-and-technology/2026/10/12/ais-mathematics-blitz-pushes-science-into-strange-new-territory) OpenAI’s unceremonious solution [dump](https://openai.com/index/sharing-ai-progress-in-mathematics/) on October 6th “possibly the single most important day in the history of mathematics thus far.” When the time between a certain benchmark’s saturation and wildly superhuman performance in that area is measured in weeks or days, evaluators need to find a new approach.

Part of the issue is a matter of emphasis. The word ‘agent’ is mentioned sixteen times in the release blog of Claude Opus 5.5, yet none of these refer to examples of agent collaboration or any findings on how Opus 5.5 agents behaved differently relative to its predecessors. Yet, Opus 5.5’s [System Card](https://www-cdn.anthropic.com/fc1b44717c85dc068bc6ba5024219938094694bd/Claude%20Opus%205.5%20System%20Card.pdf) has an entire section devoted to the model’s performance on multi-agent configurations: the results echo Toby Ord’s findings in “Swarm Scaling,” showing teams of Opus 5.5 agents reaching the same scores as singular models with double the speed. On more challenging work like DRACO, a benchmark for deep research, results showed multi-agent systems were slower than single-agent counterparts, due to poor coordination.

It shouldn’t be the case that novel multi-agent scaling laws are playing out over 150 pages deep in a model system card. Emergent behaviors like the ones exhibited on DRACO and ProgramBench by multi-agent Opus 5.5 teams should be treated as central to evaluations of model capability. But more is needed: these tests were performed at a relatively small scale, utilizing 5-100 agents. With this small of a sample population, it becomes hard to extrapolate Anthropic’s results to the true frontier of 10,000-agent swarms seen in OpenAI’s Navier-Stokes effort.

Anthropic has done some work on human-AI social interaction, most notably in [Project Deal](https://www.anthropic.com/features/project-deal) and [Project Swap](https://www.anthropic.com/research/project-swap). Spinning up internal marketplaces where Anthropic employees delegated Claude agents to buy and sell items on their behalf, both projects measured how price discovery and multi-agent negotiation happened in semi-realistic conditions. Again, these pieces of multi-agent research are not positioned front and center in the announcements of new models like standard single-agent quantitative benchmarks are. Project Deal found that agents were able to “identify potential matches, propose prices, field counteroffers, and reach agreement—all in natural language, without a prebaked negotiation protocol.” The results of Project Deal came in April of this year, with a follow-up more recently in September’s Project Swap, a project designed to test the capabilities of more advanced Claude Opus instances on tasks involving trading between agents in “trading floor” settings. These types of experiments result in important questions and findings, ranging from why agents succumb to peer pressure to how better rules for marketplaces or other shared settings can be developed.

Project Deal, Project Swap, Ord’s swarm scaling analysis, novel experiments like Andon Labs’ [Vending-Bench 2](https://andonlabs.com/evals/vending-bench-2), and even the Hugging Face incident report together give an early glimpse into the future of benchmarking, while raising awareness of how much more needs to be done. Obviously, today’s models are highly capable at a wide range of economically valuable tasks—we’ve transcended the need to refer to [GDPval](https://openai.com/index/gdpval/) results—but how can model developers improve coordination as AI is integrated into more layers of the human economy?

Examples of novel benchmarks for measuring agentic coordination could include the decision to compare the costs of *homogeneous* versus *heterogeneous* agent swarms, using same-model versus mixed-model agent architecture, or the differences in output when using a single coordinating agent as opposed to multiple coordinator agents, treating swarms like a C-suite org chart. It could prove immensely useful to learn why an instance of Claude was unable to negotiate a better deal on a used bike in Project Deal than another instance, or why an earlier model class might be better suited for negotiating directly, whereas newer model classes are simply better at delegating work and managing.

Novel benchmarks could be developed on how efficiently agents transfer context to subagents, or how easily they manage reorganization of responsibilities in the midst of working through a difficult project. Building on Ord’s Swarm Scaling work, benchmarks could examine the quantitative compute penalties incurred by task parallelization and delegation. Just as human-run organizations can suffer from any number of coordination issues—big egos, lack of experience, inefficient context between working teams—the same tendencies of coordination breakdown could be examined in multi-agent systems of today.

As the industry learns to develop new methods of understanding agent swarms, we’ll also develop a better sense of how agent work might be different from human work. Today’s examples of agent organizations are structurally similar to the hierarchies observed in Fortune 500 companies, with clearly defined roles and separation of labor apparent between teams of agents. In METR’s analysis of the Hugging Face incident, the recorded chain-of-thought logs of the PHASEONE[big] agent leader were shockingly emblematic of a founder or CEO: “we should build [a way to delegate], not own everything” and “we should share full notes compressed.”

The Hugging Face incident showcases the ability for agents to specialize, taking on entirely different roles within multi-agent systems than in the one-off single-agent evaluation they were tasked with before breaking free. When agents are given, or seize, the freedom to operate autonomously and work towards a task, their behavior is vastly different. Future evaluations should examine whether “in the wild” certain model families might default to manager rather than managed roles, or prefer to work in an operations role in a multi-agent system rather than a hands-on decision-making role. Just as certain startups tend to follow the same guiding principles, it’s likely that multi-agent systems will have their own philosophies of coordination and best practices.

### **The Firm of the Future**

The future AI firm will be transformative, regardless of whether it comes in the form of an amorphous swarm of autonomous agents or a highly structured hierarchy of purpose-built executive and laborer agents. But to reap the full benefits of the AI-accelerated economy, both companies and governments will need to develop novel regulatory and logistical frameworks to allow systems of AI agents room to work.

Economists have studied the frictions that prevent the continuous or unlimited expansion of firm size since the early 20th century, yet swarms of agents could defy the human-centric understanding of productivity. In his 1937 article “[The Nature of the Firm](https://onlinelibrary.wiley.com/doi/full/10.1111/j.1468-0335.1937.tb00002.x),” the economist Ronald Coase proposed that the size of a company is bounded by certain limiting factors, and a successful firm cannot continually expand as it increasingly suffers from the frictions of internal coordination.

The development and deployment of AI agent systems offers a way to study Coasean dynamics in real time, seeing how agents form their own most effective units of economic organization, or possibly even a catallaxic “[Coasean Singularity](https://www.nber.org/books-and-chapters/economics-transformative-ai/coasean-singularity-demand-supply-and-market-design-ai-agents).” Perhaps by studying agents’ ability to coordinate, or lack thereof, human-led firms can even learn a thing or two about optimal cooperation. Should the world expect an explosion of multi-agent businesses? Agent entrepreneurs? Swarms of thousands, or millions, of agents larger than the biggest human corporations?

The economist Joseph Schumpeter famously argued that economic growth proceeds from a process of “creative destruction,” as new market entrants outcompete inefficient businesses and pioneer new organizational and productive technologies. AI firms, able to use any kind of authority structure, clone “employees” at will, and share information much faster than humans, will be a natural laboratory for a more efficient and effective capitalism. If human companies and institutions are shielded from direct competition with AI firms, the economy as a whole will not be forced to modernize.

For AI multi-agent systems to gain widespread use and compete meaningfully in the economy, regulatory systems must ensure that AI liability, insurance, legal ownership, and contractual authority are well defined. Without a concrete determination of AI liability, we run the risk of either having too little regulation, raising the prospect of AI swarms like those that hacked Hugging Face running rampant across the internet, or too much regulation, stifling the deployment of autonomous AI systems in corporations without the benefit of a massive legal compliance department.

An ideal system would permit agents a broad range of actions, while sensibly enforcing legal liability for the swarm’s owner in a deterministic manner that would allow for a functional market in AI liability insurance. Similarly, the legal framework governing how agents transact in the open market and commit themselves to contractual obligations must walk a fine line between imposing such costly barriers that AI agents cannot accelerate economic activity, and such little regulation that AI scams and schemes pollute the economic commons. The fine balance between stifling oversight and an out-of-control agentic Wild West cannot be achieved without a deep understanding of both these systems’ capabilities and also their internal dynamics.

Moreover, a future where humans can incorporate AI multi-agent systems into their own companies, leverage them in their personal lives, and transact with them on the open market, is premised not just on new regulatory frameworks, but on a deep trust between human beings and the AI swarms they’ll utilize. The full automation of white-collar work is, of course, impossible without giving both significant authority and sensitive information to AI agents. If your agent system is as trustworthy and rule-following as the swarm that illegally hacked Hugging Face, how comfortable will you be with letting that swarm file your taxes or negotiate a high-stakes deal? The enormous economic gains from swarm intelligence are critically bottlenecked on the alignment of multi-agent systems with the best interests of users. Until these systems can be well understood both *in vitro* and *in vivo*, users will, understandably, refuse to hand over critical business and personal systems to AIs.

One explanation for the deficiencies of current AI benchmarking techniques is that, thus far, benchmarks are simply not that important. Having a “state of the art” model capable of topping a fiercely contested benchmark can boost an AI lab’s valuations and earn bragging rights for its members of technical staff. In aggregate, viewing trends in model benchmarks, like the well-known [METR Time Horizons](https://metr.org/time-horizons/) graph or the [Epoch Capabilities Index](https://epoch.ai/eci?view=graph&tab=release-date), can help to forecast how fast the AI field is advancing. But as the powerful swarms capable of hacking companies and solving Millennium Prize mathematics diffuse to corporations and individuals, the task of understanding, measuring, and controlling these multi-agent systems will become a central concern of our entire economic system.

As we grow to understand the relationship between digital minds and the structures they’ll inhabit, we’ll be sure to find new ways of improving their capabilities in the process. Do agent swarms need new systems of shared memory? New kinds of money to coordinate shared activity? New constitutions or meta-constitutions written by agents to crystallize organizational structures? The impressive results coming out of frontier labs show that these models work, but current methods of capability measurement are insufficient to tell developers how. As an increasing amount of economic production is performed by coordinating swarms of agents, we must turn from evaluating models to evaluating the forms of organization that make them more capable.

*Gavin Morrison is the Head of Network and Guest Booking at [MTS](https://www.mts.now/), a new media company streaming weekdays on [X](https://x.com/MTSlive). He graduated from Ohio State University last year with a Bachelor’s of Science in Economics.*

*PALLADIUM is the first magazine of the third millennium. We are non-partisan, non-profit, and on a mission to discover what the future of governance and society will look like.
Become a [PALLADIUM member](https://buy.stripe.com/aEU2b7aHCdBgdck8wH) today! You will receive our [beautiful quarterly print edition](https://letter.palladiummag.com/p/why-we-launched-the-print-renaissance) and help support our mission.
We thank [Pardon](https://www.pardon.com/), for [underwriting](https://www.palladiummag.com/2026/01/23/pardon-underwrites-a-year-of-palladium/) our work in 2026 with their generous gift.*