Research Radar: The Agentic State: A 20-Year Wish List, Finally Within Reach?

If you've worked in digital government long enough, you've heard this story before. Technology X will finally break down data silos. Platform Y will enable seamless cross-agency collaboration. Framework Z will deliver personalized services at scale.

At first glance, the Berlin Global Government Technology Centre's new whitepaper "The Agentic State" might read like a retrospective wish list of everything we’ve been trying—and failing—to do with technology since the early 2000s.

But AI agents are fundamentally different from previous technologies—and crucially, different from the large language models dominating current AI discussions. Unlike large language models that generate text, AI agents can perceive, reason, and act with limited human input. They don’t just inform decisions; they make them. And this whitepaper does a valuable job of showing how those capabilities could transform governance—if we can figure out how to implement them.

What This Paper Gets Right: A Systems-Level View

Rather than focusing on AI pilots or isolated use cases, the authors tackle the full stack of government operations. This comprehensive approach forces readers to think beyond chatbots and document processing to consider how autonomous systems might reshape governance itself.

• The paper treats AI agents as a systems-level shift, framing agentic AI as a structural force—capable of altering how governance itself is conceived, not just how services are delivered. 

• It offers a layered blueprint. Rather than vague aspirations, the authors organize the future of government into ten functional layers—from service delivery and compliance to leadership and procurement—each with concrete transformations and hard questions across 10 functional layers of government. 

• The paper acknowledges failure. To its credit, the paper is honest about how previous digital reforms stalled due to brittle institutions, legacy incentives, and cultural inertia. It doesn't sugarcoat the state of the state.

Why This Time Might Actually Be Different

Previous digital transformation efforts were often seen as optional efficiency improvements. But when adversaries are already weaponizing AI and citizens can access sophisticated AI services from private companies, government adoption becomes a matter of legitimacy and security rather than just optimization.

Several factors distinguish current AI capabilities from previous waves of digital transformation promises:

Unlike previous automation efforts that required perfect data integration and standardized processes, AI agents can interpret unstructured documents, work across inconsistent databases, and apply judgment to edge cases. This means governments don't need to achieve organizational perfection before seeing benefits.

The authors highlight a shift toward outcome-based procurement where governments pay for results rather than tools or staff hours. If AI providers take responsibility for delivering measurable outcomes, this could bypass traditional procurement bottlenecks and risk-averse bureaucratic cultures.

Previous digital transformation efforts were often seen as optional efficiency improvements. But when adversaries are already weaponizing AI and citizens can access sophisticated AI services from private companies, government adoption becomes a matter of legitimacy and security rather than just optimization.

The paper argues, convincingly, that agents can deliver hyper-personalized services and real-time compliance at near-zero marginal cost, something that was never possible in the past.

Where Version 2 Should Head: From Vision to Viability

If Version 1 was about establishing a compelling vision, Version 2 should focus on how we get there. Here are six suggestions:

1. Define the threshold between LLMs and agents.

The paper is enthusiastic about agents but vague about what actually makes them distinct. A typology of agentic capabilities—planning, memory, autonomy, coordination—would help governments evaluate what's safe and feasible today versus what remains speculative.

2. Lay out a realistic transition pathway.

What does a “crawl–walk–run” adoption timeline look like? What foundational capabilities need to be in place (e.g., identity systems, metadata standards) before agents can be effective? Show how a government moves from current state to agentic operations in realistic steps.

3. Address legitimacy and trust.

Agentic AI in public service isn’t just a technical or economic shift—it’s a democratic one. The paper needs a deeper dive into accountability, transparency, and how to preserve public legitimacy when machines are doing the governing. Expand the discussion of government as an AI platform provider. How do nations build public agentic capabilities that serve democratic values while remaining competitive? What does it mean to treat AI models, datasets, and agent frameworks as public infrastructure? Connect this conversation to the burgeoning debate about Public AI. We’ll be having some of that conversation in Princeton later this month.

4. Name the political economy.

Who stands to gain or lose from an agentic state? What happens to professional roles, labor unions, procurement incumbents, or regulators? The power dynamics matter—and will shape what’s iplementable.

5. Focus on institutional design

The fundamental challenge isn't technical—it's that the institutions, legal frameworks, and cultural norms governing public administration weren't designed for autonomous systems making real-time decisions. We need to ask, as the Institutional Architecture Lab at UCL does, what kind of organizations do we need if we want to realize the ethical adoption of AI agents in governance?  How do existing democratic institutions like legislatures, courts, and oversight bodies need to evolve? 

6. Confront the capacity gap head-on.

How do governments with fragile data infrastructure, talent shortages, and risk-averse cultures actually deploy these systems? Case studies or early pilots—warts and all—would add credibility. Here is an example of how we’ve been using AI agents in New Jersey. Also at InnovateUS, we are planning a free, skill-building workshop series for the fall on Agentic AI and policy making this fall and would love to collaborate.

Let's make sure we build an agentic future that serves not just efficiency, but equity, legitimacy, and public value.

Moving Beyond the Wishlist

The Agentic State is a bold, thoughtful paper that confronts decades of stalled reform and asks: what if we could finally build the public institutions we’ve long imagined?

The authors are right: this transformation is likely inevitable. The question is whether governments will lead it through deliberate institutional innovation—or be dragged into it by technological and geopolitical forces.

Version 2.0 is the moment to shift from describing the destination to charting the route. Let's make sure we build an agentic future that serves not just efficiency, but equity, legitimacy, and public value.

Read the full whitepaper here. To read more about the opportunities posed by AI to enable more responsive democratic institutions, click here.