A Government of Agents: The Experimental Logics of Artificial Intelligence

Across climate science, logistics, pandemic response, and military operations, a new class of AI platforms is emerging that treats the planet as a laboratory and politics as a matter of automated bureaucracy. The European Union is building a “digital twin” of the entire Earth—a system it describes as “a highly-accurate digital model of the Earth to model, monitor and simulate natural phenomena, hazards and the related human activities.”1 Nvidia sells AI weather prediction as a cloud service. Google has launched a suite of applications—GenCast, FireCast, and others—for environmental management and disaster modeling. Palantir offers platforms for simulation, modeling, and twinning, marketed as “AI-powered automation for every decision,” from food security to battlefield operations.2 Chinese corporations are building their own planetary-scale platforms, such as Alibaba’s City Brain for urban management and Huawei’s Safe City technologies, which are deployed across China, Africa, and Southeast Asia promising “resilience” in the face of environmental and security crises.3 

All of these platforms share a common logic. In Palantir’s own language, they are designed to “implement, manage, test, and evaluate advanced models on top of live data for improved forecasting—incorporating outcomes as data for continuous improvement.”4 The key word here is “test”: these systems are imagined as doppelganger worlds, cybernetically linked to the real one, capable (in theory) of reallocating decisions and affecting outcomes in real time through predictive modeling, automated action, and increasingly autonomous decision-making. Most importantly, they are imagined to be constantly experimenting with the future—“continuously improving” the systems they manage. Technology, in this vision, bequeaths evolutionary capacities to technical systems. The imaginary of agentic decision-making is thus integrally linked to the idea of constant testing and experimentation—the only mode by which machines learn. 

This transformation can be given a name: agentic governance. The term names the imagined reallocation of human agency and decision-making to automated systems—systems authorized, and imagined, as capable of autonomous decision and emergent behavior—on the assumption that planetary systems are permanently in crisis and too complex for human cognition alone, requiring substantial delegation to vast computing infrastructures. Agentic governance is an emerging rationality—a way of thinking about power and agency being built, platform by platform, into the operating systems of the twenty-first century.

Contemporary forms of AI-driven decision-making unsettle the relationships between knowledge production, political deliberation, and democratic decision-making. In the postwar period, scientific legitimacy was increasingly associated with falsifiability: a claim counted as scientific only insofar as it could, in principle, be refuted.5 What, then, becomes of scientific and democratic accountability when inquiry no longer begins with an explicit hypothesis? This question compounds longstanding concerns that surveillance capitalism, automated decision-making, and diminishing confidence in law’s capacity to govern complex technologies are eroding the free flow of information, privacy, and public debate on which democratic government depends. The emergence of agentic governance goes further still, challenging the concept of human agency at the core of liberal governance. Where democratic legitimacy rests on deliberation before action, governance by experiment acts first and assesses afterward. Feminist science studies scholars have long noted that liberalism is a fantasy, and one long challenged by cybernetics. Donna Haraway’s account of the cyborg exposed the instability of the boundary between human and machine.6 Yet the scale, integration, and institutional organization of contemporary AI infrastructures arguably represent a qualitative shift, demanding new conceptual resources for analysis and governance.

Experimentation and Computing

The concept of experimentation lies at the foundation of agentic governance, the logic of which Google’s chief economist Hal Varian clearly articulated a decade ago. Since correlation is not the same as causation, he argued, the solution is experiments; “ideally these would be carried on continuously” using the infrastructure of the web. In a parallel paper from 2010, Varian championed “natural experiments,” leveraging real-world events like natural disasters, policy changes, and geopolitical disruptions to test cause-and-effect relationships on whole populations, arguing that such practices should guide economics to becoming a “hard science.” Public health served as his model.7 In 2021, the Nobel Prize in Economics was awarded to David Card, Joshua Angrist, and Guido Imbens for precisely such methodological contributions. The transformation of shock or catastrophe into a repeatable and explicitly experimental setup already signifies a new stage in the history of disaster capitalism. In this logic, the experimentation itself becomes a site of assetization, both an epistemology and a business technology. 

The Covid-19 pandemic supercharged the logic of “natural”—that is, planetary-scale—experimentation. The integration of geolocation data gathered from cell phones and smart infrastructures transformed the modeling of global disease spread, turning a public-health catastrophe into a massive natural experiment in real time. At the heart of these ongoing experiments is an infrastructural imaginary of producing digitized twins of our planet—what the computer scientist David Gelernter prophetically described in 1991 as “mirror worlds.”8 The transformation of emergencies and planetary-scale events into continuous experiments is historically significant, and takes on new importance when situated within the broader changes in government and economy now happening with AI.

Planetary Test-Bedding

There is a history to this present. The Trinity test of the atomic bomb in July 1945 was a planetary experiment in the most literal sense: physicists debated whether the detonation might ignite Earth’s atmosphere and oceans. It did not. But the subsequent use of nuclear weapons—in Japan, and in tests by the United States, the Soviet Union, France, the United Kingdom, and China through the 1960s—scattered radioactive fallout that increased atmospheric carbon-14 across every carbon cycle on the planet.9 This became a permanent geological signature, enabling entirely new forms of experimentation such as geological dating, ecological studies, and forensic testing. Some scientists have proposed it as the physical marker of the Anthropocene. 

The nuclear age also generated the institutional architecture for managing planetary-scale risk through simulation, and inaugurated a discourse of fallibility in relation to human decision-making. The Cold War produced new forms of technocratic expertise—scenario planning, simulation, game theory, operations research—designed to govern emergencies at global scale.10 A new class of experts arose whose job was to think the unthinkable and to build systems that could act when humans could not think fast enough.

Some of the first computational models of the entire Earth system were born in this context. The Limits to Growth report, published in 1972 by the Club of Rome, is often cited as the beginning of the modern environmental movement. Assuming that human politicians could neither grasp nor manage the complexities of the environmental change humans had created, a group of computer scientists and demographers decided it was time for researchers armed with quantitative and computational methods to demonstrate and explain the implications of human behavior and policy to the world.11  

Using a computer model called World3—developed from computer scientist Jay Forrester’s pioneering work on system dynamics, which had migrated from modeling industrial supply chains to urban planning and finally to the planetary scale—researchers at MIT simulated the planet’s future across five variables: population, industrialization, pollution, food production, and resource depletion. Their conclusion was dire: the global system would collapse within a century if growth trends continued. But collapse was not inevitable. “It is possible,” the authors urged, “to alter these growth trends and to establish a condition of ecological and economic stability that is sustainable far into the future.”12 

Limits to Growth was built on a particular vision of knowledge. The model was deductive: it started from rules derived from observed systems and projected forward. It assumed the Earth was representable, amenable to computation, and ultimately controllable through planning and scientific expertise. The report posed the planet as being in crisis while simultaneously insisting it was scientifically and technically controllable. Agents in the simulation could not learn or evolve. Human behavior was a “problem.” The planet, to use Buckminster Fuller’s famous adage, was imagined as a spaceship—enclosed, manageable, awaiting the right engineers.13 The Cold War championed a model of rationality that was algorithmic, repeatable, and quantitative, closely correlated with expertise grounded in cybernetics, game theory, and operations research. The world was a problem; knowledge was a solution. And the tools of simulation and computation would deliver it.14 

Knowledge and Power

Agentic governance through experimentation marks a new relation between knowledge and power. If the Cold War inaugurated a logic of technocracy ruled by expertise, contemporary AI platforms have another, more operational logic. Nvidia’s CEO Jensen Huang described his Omniverse platform—the substrate for Earth-2, “a fully open software stack that accelerates every stage of forecasting, from data processing to high-resolution visualization”—as a tool that would automate inference and embed it in the environment through cloud computing and the Internet of Things.15 This automated inference, he claimed, enables “never-seen-before” predictions.16 This phrase, “never-seen-before,” signals a fundamental transformation in what counts as knowledge when decisions can be delegated to inference.

Inference, in AI, is when a trained model applies its learning to data it has never encountered before. It is what happens when you ask a chatbot to summarize an article or when a weather model predicts tomorrow’s storm. But the conclusion is drawn without a causal or explanatory model. The system does not understand why the storm will happen. It identifies statistical patterns in historical data and extrapolates them forward. It correlates; it does not explain. Researchers call this “generativity”—the production of new outputs from past data without establishing causation. Neural networks cannot determine that A causes B. There is a model, but it is not a model of the world. It is a model of statistical relationships between data points—a model of the architecture of machine learning itself.17 

This represents a clean break from the tradition Limits to Growth embodied. In the most advanced current efforts at weather modeling—by Google (GenCast), Nvidia (Earth-2), and the EU’s Destination Earth program—the frontier is what researchers call “non-numeric” weather forecasting. Rather than modeling the underlying physics of weather, which requires linking expertise from multiple scientific fields, these AI systems analyze historical weather images and environmental datasets through statistical learning alone. They do not simulate the fluid dynamics of the atmosphere. They learn from satellite imagery what, for example, a storm looks like, and use that pattern to forecast future storms; or they learn which combinations of variables—air pressure, wind, and other indicators—correlate with certain weather formations. In short, these models take the data historically gathered by existing Earth-observation systems, find patterns, and extrapolate them into the future. Other techniques—diffusion methods, ensemble data assimilation, and related approximations of Monte Carlo simulation—add statistical variation to the predictions, but none of them are grounded in the physics of weather.18 The results are often remarkably accurate for short-term forecasts—in some cases surpassing the physics-based models they are beginning to supplement and, as some believe, may eventually replace. What is lost is explanatory or exploratory ideas of science. The world is always already known in AI modeling, and the future is always derived from the past—even if not linearly or causally. 

The implications of this are profound. When weather can be predicted without climate models, when correlation appears to suffice without causation, something fundamental can shift in the political landscape of climate science. AI-driven prediction offers an arrangement that is, whether intentionally or not, ideologically convenient: accurate forecasts without the politically inconvenient attribution of cause. You can prepare for the hurricane without acknowledging why hurricanes are getting worse. In a context where anthropogenic climate change remains bitterly contested—and government climate and environmental research and monitoring agencies are being cut around the world—non-causal prediction functions as a way of managing consequences without confronting origins. It also permits a new group of actors—mainly corporations and AI platforms—to play increasingly important roles in environmental management at planetary scales.  Whether this represents a genuine advance in scientific method or a convenient depoliticization of climate knowledge—or both—will shape environmental policy for decades. 

Palantir website, ➝. Screenshots taken January 4, 2026.

Automating Decision Making

If the epistemic shift brought by AI modeling is about what counts as knowledge, the operational shift is about what counts as a decision—who will be making decisions, and how they will be held accountable. At the heart of digital twinning is the idea that the future can be managed, but not through official policy directives or plans. Instead, management happens through automated incremental testing—or, in AI language, “learning.” The origins of the term “twin” reveal this aspiration. NASA engineers working on the Apollo 8 mission in the 1960s coined it to describe a cybernetic vision of simulation in which a virtual model would be tied to its physical counterpart through continuous loops of real-time data exchange. What differentiated digital twins from older simulation technologies, NASA leaders later explained, was scale, ordinality, and the non-deterministic nature of the models.19 Twins could address phenomena impossible to trial within a laboratory, like planetary modeling, stochastic futures, or irreducible complexity. They were imagined as a technology built for the unforeseen accident or disaster, and they are credited with helping rescue the failed Apollo 13 mission. 

In our present, this logic of automated adaptation has been extended from space capsules (such as Fuller’s Spaceship Earth) to the entire Earth system. No company embodies this transformation more aggressively than Palantir. Co-founded by Peter Thiel, currently led by CEO Alexander Karp, and started with a CIA venture fund in the wake of 9/11, Palantir is now valued at roughly $440 billion. The United States government is its largest client; in 2025, the company received the largest software contract in the history of the Department of Defense, valued at $10 billion. It provides software to the World Food Programme, the Ukrainian military, the UK’s National Health Service, and a sprawling roster of other top-level government, non-profit, and corporate organizations ranging from police forces and militaries to medical institutions. Civil rights advocates and scholars have long scrutinized its role in predictive policing, border surveillance, and its work with US Immigration and Customs Enforcement. More recently, questions have been raised about its possible involvement in the Gaza conflict.20 

I recently watched one of Palantir’s demo videos from 2021 for its Gotham platform—named, with characteristic grandiosity, after Batman’s city, whose superheroic capacities to forestall crime Palantir apparently aspires to emulate.21 The scenario takes place in the Taiwan Strait. I am looking over the shoulder of a military analyst. AI systems reading Japanese satellite imagery identify fishing boats lashed together in suspicious formations. Simultaneously, a Chinese destroyer departs a port and goes dark. Boxes flash across the screen at impossible speed as the system searches through vast data fields—drone imagery, ocean sensors, ship data, feeds from multiple intelligence agencies and navies. Decisions need to be made. The system rapidly triangulates data from governments, weather and environmental sensors, field troops, and the US and Japanese navies, running real-time simulations and testing possible outcomes. Multiple courses of action are plotted and displayed; new parameters are added, more simulations produced. In only ninety seconds, the system reallocates agency among military personnel dispersed across the planet, generating a coordinated decision—launching drones, repositioning ships—that averts the threat. The world’s semiconductor supply chain is secured. I can continue browsing the internet, safe in the knowledge that there will be ever more chips for AI training.22 

Left: Diagram from Palantir Technologies Inc, "Creating Data in a Data Store Using a Dynamic Ontology," US Patent 7,962,495 B2, November 20, 2006. Right: Diagram of Palantir's "Ontology" system, a central feature of all its platforms. Screenshot from Palantir Foundry overview (2024).

This marketing video reveals the platform’s logic with startling clarity. One feature that stands out is prompting: analysts do not receive a complete picture of the situation. They receive AI-recognized features and interact with the system through iterative queries, each generating new “courses of action.” The analyst is not interpreting the world; the analyst is steering an agentic machine that interprets the world. As scholars Louise Amoore and Lucy Suchman have argued, this kind of automated warfare interpolates the human operator into a continuous testing relationship with the system—one that attunes the human to the possibility space the machine allows, rather than the reverse.23 Rather than visualizing the entire battlefield, the system coordinates and automates cascading micro-decisions that cumulatively generate large, emergent outcomes.  Every event becomes a scenario decomposed into relational data points in a series of ongoing micro-decisions delegated across humans and machines that evade any need to situate, contextualize, or comprehend the event historically or politically. Palantir calls this an "ontology": a continuously revised digital twin of the operating theater in which representation and action are fused, so that agents are never fixed entities but are perpetually redefined by the decisions taken through the system itself. The system replaces comprehension with automated coordination—and calls this intelligence.

Whether Palantir’s technology actually delivers on these promises is an open question. Skeptics note that the company’s meteoric rise in government contracts may owe as much to the political connections of its founders as to technical superiority.24 But the vision itself has force. It is being funded at planetary scale, and it is shaping how institutions from the Pentagon to the National Health Service imagine the future of automated decision-making.25  

DestinE “Data Lake.” Image from Cezary Mazurek and Maciej Stroinski, “Technology Pillars for Digital Transformation of Cities Based on Open Software Architecture for End2end Data Streaming,” in Proceedings of the 55th Hawaii International Conference on System Sciences (2022).

Territory

If these platforms aspire to planetary reach, the political geography they produce is anything but unified. For example, Destination Earth (DestinE) is embedded in the European Green Deal and the European Digital Strategy—it in theory serves a political bloc, not the globe. Yet modeling climate impacts on European food security requires global weather data. DestinE therefore networks EU data with data gathered elsewhere on the planet within a “Data Lake.” The result is a planetary data-gathering and decision-automation apparatus harnessed to a regional political project. Later phases of the project aim to add capacities for optimizing urban traffic, tracking biodiversity, quality-of-life monitoring, and “smart” farming—all within the EU’s jurisdictional frame. This territorial vision maps onto neither Westphalian nation-states nor Cold War superpowers, but onto competing data spaces and governance blocs that overlap and contest one another while collectively claiming responsiveness to Earth-system crises. In the EU’s case, this is at least correlated with a democratically elected governing body and abides by laws interested in transparency and privacy. But this is not the case everywhere AI is being developed to manage large infrastructures at scale. Borders, wars, prisons, and logistical systems are the sites where digital twins are most commonly found, and where the politics of these platforms is most clearly made visible globally.26 

The EU's digital twin, which manages planetary risk by simulating future climate events, is only one instance of a larger experimental relationship to AI—one aimed at managing security and perhaps even democratic oversight itself as a risk. The United States pursues an alternative but similarly “agentic” logic in its policies guiding AI platforms. The second Trump administration’s “America’s AI Action Plan” centers on American infrastructure—server farms, energy grids, chip production—while banning state-level AI regulation and recommending regulatory sandboxes outside formal oversight.27 The Chinese have any number of similar policies encouraging global infrastructure development, including the Digital Silk Road project that exports integrated digital infrastructure—including 5G networks, cloud computing, and intelligent surveillance systems—to partner countries worldwide.28 While facilitating this expansion of surveillance capabilities, the initiative simultaneously deploys big Earth data and remote sensing networks to conduct large-scale environmental and climate monitoring in support of regional sustainable development goals.29 

There have long been infrastructures for experiment. For example, nineteenth-century colonial projects such as the Magnetic Crusade demanded networks of observatories across the planet. Development projects in the second half of the twentieth century as part of the Green Revolution used test fields spanning the globe and were managed by international agencies such as USAID, the World Food Programme, and the Ford and Rockefeller foundations. And weather and climate modeling has long used extra-planetary sensory infrastructures such as satellite systems. However, the contemporary acceleration and intensification of these historical systems marks, I argue, an epistemic change.30 This change has been anticipated. “Surveillance” and “platform capitalism” already name some of the modes of assetization emerging.31 But the contemporary push for agentic bureaucracy and automated decision-making at scale, and its shift in how the world is represented and rendered knowable and actionable through experiment, is new.32  

Experimental Life

Twenty years ago, Bruno Latour made an observation that has only grown in significance over time: we are all part of experiments. The climate crisis, genetic engineering, pandemics—these are not problems being studied in controlled conditions by credentialed experts. They are, Latour argued, “collective experiments” that “have spilled over the strict confines of the laboratories” drawing every one of us into trials conducted “at scale one”—that is, planetary scale—"and in real time.”33 The Earth itself has become a test bed. If this once sounded like provocation, it now reads more like a mission statement for Silicon Valley, where the notion that the world exists to be tested upon has become foundational. 

The erosion of expertise, the infrastructuralization of experiment, and the transformations in knowledge and time also refract a broader understanding of human relationships to the Earth. The core ideology of agentic governance is that it cannot itself be governed (even critics forward this idea): automated systems make decisions at a speed and scale where older forms of policy, law, and regulation can barely operate.34  

The pervasive narrative that democratic institutions are too slow to comprehend, let alone govern, rapidly evolving artificial intelligence has fueled the global proliferation of the “regulatory sandbox.” The sandbox, designed as an official mechanism for managing emerging technologies, creates an environment in which regulators suspend the ordinary application of rules for a designated firm, product, or trial—within set limits and timeframes—in order to observe and learn. Rooted in the assumption that AI exceeds both human control and the deliberative pace of lawmakers, these designated spaces exist to test and release technology in advance of formal legal frameworks. Yet the containment the sandbox promises is far from assured.35

The ungovernability of the technosphere is not hyperbole or theory. Worldwide, the sandbox has become a dominant regulatory paradigm. Nations like Singapore have established themselves as leaders in this proactive approach to regulation. In the United States, recent presidential decrees by the Trump administration advocated for sandboxing while restricting state-level abilities to regulate AI, arguing that traditional legal oversight stifles innovation because democratic governments are theoretically too slow and incapable of timely deliberation.36 The European Union is similarly experimenting with sandboxes, operating under the assumption that standard regulatory oversight of AI is simply no longer viable.37 

Ironically, both state governments and critical scientists frequently concur on this understanding of agentic governance, agreeing that the crisis of technological velocity can only be negotiated through constant automation, experimentation, and learning, i.e. technology itself. However, this continuous experimentation fundamentally inverts the traditional relationship between democratic deliberation and technological deployment.38 Ideally, democratic politics dictates that consequential decisions must be preceded by public debate, expert consultation, and legislative action. The logic of sandboxing reverses this sequence: automated systems are deployed first, with their societal effects assessed retroactively through ongoing “learning.” 

However, DestinE, and digital twins of the Earth more generally, might counter this assumed impossibility of democratically governing agentic systems by producing new images of governance. Digital twins of the Earth presume that human policy expertise can neither grasp the enormous volume of information generated by global natural systems such as weather nor adjust policy at the necessary speed.39 In this sense, such systems assume only agentic governance can manage the planetary scale of climate crisis and technical change. The political aspiration of DestinE, however, is to maintain a human dimension by suggesting that through sufficiently large collectives (the entire EU and all of its partners globally), humans can introduce some limited forms of management and governance to machine systems, even if only through offering data and administration. DestinE presumes that such governance is possible because agentic infrastructures are both Earth systems and infrastructures for enabling fine-grained digital twins of the Earth. Insofar as DestinE aims at a mode of governance that is necessarily partially automated by AI, such governance is an agentic combination of human judgment and machine decision. But its infrastructures, knowledges, funding, and economy are clearly organized and deliberated upon by national bodies. As a system, it remains under the guidance of a presumably democratic polity. 

Ulrich Beck already suggested “reflexive modernization” as a possible response to the risks produced by modernity, whereby the instruments and technologies that produce risk might also be used to monitor, reflect on, and perhaps mitigate those risks.40 For not all experiments are the same, and recognizing this heterogeneity is essential for political response. The question of experiment as a public forum has never been more critical, and the structure of the institutions running these trials is of the highest urgency. DestinE’s aspiration to make data and results available to the public represents a fundamentally different political vision from Palantir's proprietary platforms or Nvidia’s corporate research. The difference is not merely technical but concerns the very possibility of democratic accountability in an age where agentic governance is being widely propagated. The answer appears to lie in organizations, institutions, and infrastructures—in how we imagine and therefore build them.