When modern governments took shape, the dominant idea of what it meant to understand something was to decompose it. Break a problem into its constituent parts, assign each part to a specialist, and govern the whole by governing its pieces. Behind this organizational logic stood an entire worldview, one built on reductionism and determinism, on the belief that complex phenomena could be mastered by isolating their components and studying them independently. The silo model of government is that worldview made institutional. Every ministry, every agency, every sectoral database that cannot share information with its neighbors is reductionism in organizational form, a structure whose implicit premise is that governing the pieces exhausts the government of the whole. This premise was never stated explicitly, because it never needed to be. It operated as a background assumption so deeply embedded in the design that the design itself made it invisible.
The model worked extraordinarily well for problems that are in fact decomposable, stable, and linear, which is to say, for the class of problems that governments were originally designed to handle. Tax collection, vehicle registration, the issuance of documents and licenses, the administration of pension records, these are weakly coupled phenomena where the analytical procedure delivers exactly what it promises. But the problems that define our era are not those. Poverty is not a housing problem with health and education side effects. Climate adaptation is not an environmental policy with economic externalities. Youth unemployment in a digital economy is not a labor market anomaly that the labor ministry can resolve on its own. These are deeply interdependent phenomena, what complexity theorists call “wicked problems”, in which the boundary between domains is the artifact of administrative design rather than a property of the underlying reality, and in which intervening on one variable reshapes all the others in ways the original intervention did not anticipate.
The silo state is not failing because it is underfunded or poorly managed. It is failing because it was designed for a different kind of problem, and the most revealing symptom of this failure is what happens when governments try to fix it. They reach for the same toolbox. Another coordinating committee, another interministerial task force, another portal that announces integration without achieving it, another strategy document that maps the same territory the last one mapped. The reform inherits the logic of the thing it is trying to reform. Decades of joined-up government initiatives, whole-of-government programs, and digital transformation strategies have produced abundant evidence of this pattern. The infrastructure of decomposition is not reformed by adding coordination on top of it. The foundation itself has to change.
A Different Foundation
The sciences of complexity have been pointing toward an alternative for decades, and the core thesis is straightforward: AI agents and digital public infrastructure (DPI), organized around principles of systems thinking, enable governments to transcend the silo model and work in an orchestrated and integrated way. This claim rests on three concrete ideas that complexity science makes precise. Interdependence means recognizing that the phenomena governments govern do not respect institutional boundaries, and that the infrastructure of government must reflect this rather than suppress it. Emergence means that when the right capabilities are coupled through shared data flows and agentic workflow orchestration, the system acquires properties that no individual institution could produce alone, a capacity to perceive and act on a citizen's situation as a whole rather than as a collection of sectoral cases. Adaptation means that feedback, measurement, and learning are built into the architecture rather than bolted on as afterthoughts, so that the institution improves through operation rather than only through periodic reform.
The result is what we call complex institutions, organizations that bring together diverse data flows to support system-level policies through agentic workflow orchestration. These are not larger bureaucracies, not more powerful ministries, not better-staffed coordination offices. They are a qualitatively different kind of institution, one whose architecture treats the relationships between domains as first-class objects of governance rather than as problems to be minimized or assigned to a committee.
Most governments today are not standing still. They are adopting AI strategies, publishing digital government roadmaps, and establishing frameworks for responsible AI use. However, there is a significant and widely documented gap between the policies governments declare and the services citizens actually experience, and the source of this gap is consistent across contexts. The shared infrastructure required to deliver integrated, intelligent services at scale simply does not get built. Strategies are adopted and services remain fragmented. Roadmaps are published and databases remain siloed. The gap between policy and reality is not a failure of vision, but a failure of implementation, and it persists because the thing that needs to be built, a common technical substrate that makes orchestrated government actually executable, is neither glamorous nor politically legible in the way that a strategy document is.
From Digital Government to Agentic Government
For a generation, the aspiration of digital government was to move services from paper to screen. With AI, the chatbot became the symbol of this era, an interface that could explain the requirements for a permit, describe the steps in a process, and direct citizens to the relevant office. It could not do anything. The shift to agentic government is a shift from answering questions to executing services, and it changes the fundamental nature of what a government can be. An AI agent does not explain the steps to obtain a birth certificate. It verifies your identity against official records, retrieves the required documentation from the civil registry, checks the applicable rules encoded as machine-readable law, prepares the application, and asks for your authorization to submit it. The interaction moves from informational to transactional, from a state that describes its services to one that performs them, from a model in which the citizen must navigate the complexity of government to one in which the government navigates on behalf of the citizen.
Over the past several years, the Center for Public Intelligence (CPI) has been building the infrastructure this vision requires. The result is GovOS, an operating system for agentic government, designed as an ecosystem of systems organized around a shared core. To understand what it does, consider what happens when a citizen requests a birth certificate through a single interaction. Behind that request, GovOS orchestrates every capability the transaction requires. Identity is verified against official records. Interoperability layers query the civil registry. Documents are retrieved and validated. Payment is processed. Notification is sent. The service is completed. One citizen interaction. Multiple government systems. No visible complexity. This is the architecture of the complex institution made concrete, and it is already in production.
This shift is made possible by public AI agents that can transact services across the full spectrum of government functions. Such an agent operates through natural language, understanding voice and text, delivering services through any channel, from chat and SMS to phone, web, and app. It provides proactive rather than reactive service, anticipating citizen needs based on life events rather than waiting for applications to be filed. It is omnichannel by design, because a state that can only be reached through a smartphone has not solved the access problem but digitized it. The aspiration is a system so capable that the complexity of interagency coordination becomes invisible to the citizen who experiences it.
Agent ID: The Governance of Complex Institutions
The most consequential architectural innovation in this approach is also the most underappreciated. It concerns not what AI agents can do, but how they are governed. The insight is simple and its implications are large: AI agents deployed by the state should have the same institutional identity infrastructure as human civil servants.
Agent ID gives each AI agent deployed by a government a formal institutional identity: a name, a unique identifier, a role, a status, a defined set of permissions, and a defined scope of access to government systems. Just as the human identity layer of government gives a civil servant the permissions required to perform their role and no others, Agent ID does the same for AI agents. No agent executes without identity and authorization. Every action is attributable to a specific agent operating under a specific legal authorization within a specific workflow. Every permission is revocable. The chain from agent to identity to permissions to execution to audit is the condition of deployment.
This matters because it answers the AI governance question at the architectural level rather than the policy level. Declarations that AI will be used responsibly are not governance. Governance is what happens when accountability is made structural, when the question of who authorized this action and on what legal basis can be answered for every action an AI agent takes, not as an ex post audit but as a property of the execution itself. An AI agent that verified identity, retrieved documents, and initiated a payment did so with a specific identity, under a specific legal authorization, in a workflow that recorded every step.
What This Requires
None of this happens by declaring a policy or publishing a strategy. It requires building shared identity infrastructure, interoperability layers, and clear legal purpose, a governance model for AI agents that makes every action attributable and every permission revocable, and open models and datasets that any public institution can use without dependence on a single vendor. It requires the political will to invest in shared infrastructure before the coordination problem becomes a crisis, and the institutional courage to govern AI by design rather than by declaration. This is already being built. AI is already transforming government. The question is whether that transformation will be organized as public intelligence, governed democratically and accessible to all, or as private intelligence, concentrated in the hands of those who built it first and structured to serve those who can pay for it. The government was not built for the world we live in. Building the one that is, openly, accountably, and with the sovereignty of every people in mind, is the work in front of us.