Three Paths for the Agentic State, 2026–2030
Foresight · Synthocracy Institute · 5 July 2026
This is foresight, not forecast. The scenarios below are structured, deliberately divergent pictures of paths societies could take. They are built from trends we can observe today, but they are not predictions, and nothing here is offered as settled fact. Their purpose is to sharpen a choice while it is still open — to show what different decisions would make ordinary, and what they would make impossible. We hold this work to its own honest standard: it informs our research, and it never poses as established science.
By 2030 much of the state’s routine business — who gets a benefit, a licence, an inspection, a place in a queue, a flag for review — will be run not by officials advised by software but by AI agents that act. The open question is not whether this happens. It is what kind of decision order it produces. This paper sketches three: a state that stays legible, a state that drifts into ceremony, and a state that seals itself. Most real jurisdictions will end up as a patchwork of all three. The value is in seeing the pure types clearly enough to steer toward one and away from the others.
What we are watching (the 2026 baseline)
The scenarios are extrapolations from things already true in 2026, not from imagination.
Adoption is moving from advice to action. Industry analysts project that at least 80% of governments will deploy AI agents to automate routine decision-making by 2028, and government surveys already describe the agentic era as arrived rather than approaching. The first governance templates built specifically for agents exist — Singapore’s IMDA framework (January 2026), the U.S. NIST agent-standards work (February 2026) — but they concede that the hardest parts, dynamic agent identity and delegation chains, remain unsolved. The binding legal backstop in Europe for exactly the high-stakes public domains — justice, public administration, essential services — was deferred by the Digital Omnibus to December 2027, arriving after the agents. In the United States, the direction is toward preemption and deregulation, with the accountability action migrating to states: Colorado’s pivot to a narrower automated-decision statute keeps only notice, an adverse-outcome explanation, and a right to human review. And two older cases — Australia’s Robodebt and the Dutch childcare-benefit scandal — stand as proof that automated decision systems can meet every compliance requirement on paper and still cause mass harm.
Read together, these are the raw materials of very different futures. What separates the paths is not the technology. It is whether two things keep pace with the agents.
The variable that splits the paths
Does the record keep pace with the action? When an agent acts, is there — before it runs — a record of its identity, its authority, its delegation chain, and a real human boundary? Or does adoption outrun the records that would make it accountable? Call this legibility.
Where does control settle? Does decision-making power stay distributed and contestable — visible to citizens, open to challenge, reversible — or does it concentrate into public-private stacks of vendors, models, and cloud that no outside party can inspect? Call this concentration.
The three paths are what you get when these variables break in different directions.
Path One — The Accountable State
Legibility keeps pace. Power moves into the machine, but stays visible.
In this path, governments treat the arrival of agents as an admissibility problem and build the record before the agent acts. Agent registries become as ordinary as vehicle registries. Every consequential public action carries an identity, a recorded autonomy level, and a delegation trail that can be reconstructed — which agent acted under whose authority. The citizen-facing side of Colorado’s statute generalises: when an AI-run decision goes against you, you are told, you are given the actual reason, and you can reach a human with the standing to reverse it. Human oversight is tiered honestly — automated where actions are reversible and low-stakes, a named person signing where they are not — and override rates are audited so that “human review” cannot quietly become a rubber stamp.
Here AI genuinely co-decides, and that is not a scandal, because the co-deciding is on the record. Synthocracy exists — power has passed into the machine — but it is governed: inspectable, challengeable, reversible. This is the most demanding path. It costs money, slows some deployments, and requires institutions to admit publicly where AI is co-deciding rather than merely assisting.
Signposts it is emerging: agent registries and identity requirements written into procurement; delegation-chain audit trails mandated for public-sector agents; adverse-decision explanation rights spreading beyond one or two states; override-rate reporting appearing in public audits.
Path Two — The Ceremonial State
Legibility falls behind. Power moves into the machine, and no one can quite locate who decided.
This is the path of least resistance, and the most dangerous precisely because it looks fine from the outside. The agents run the workflow. Humans remain “in the loop,” clicking approve. Dashboards multiply. Every compliance box is ticked. But the boundary is ceremonial: by the time the human clicks, the agent has already determined the case — set the default, ranked the queue, written the summary, prepared the route — and the click launders that determination into a human decision on paper.
The power to act has migrated to the machine while the authority to answer for it stays nominally human. When something goes wrong, it goes wrong the way Robodebt and the Dutch childcare scandal went wrong — at scale, at speed, quietly, and in full compliance. Adverse-decision explanations exist, but they are boilerplate. Audit trails exist, but they record clicks, not reasoning. The state can always show that a human was present; it can never quite show that a human decided.
This is synthocracy as drift: no one seizes power, it simply relocates, unannounced, and the institutions keep issuing the paperwork of accountability over a decision layer that has moved beyond their reach. It is the default outcome if nothing is built to prevent it — which is why naming it matters.
Signposts it is emerging: “human review” approval rates so high, and so fast, that genuine review is statistically implausible; adverse-decision notices that are visibly templated; audit logs that capture confirmations but not the alternatives suppressed or the reasons given; a widening gap between systems labelled “AI-assisted” and systems that are, in operation, co-deciding.
Path Three — The Sealed State
Control concentrates and seals. Power moves into the machine, and behind a wall.
In this path the pressures of national security and efficiency push the state’s agentic decision layer into a small number of public-private stacks whose reasoning is classified, proprietary, or both. The logic first seen in 2026, when a government exercised an admit-hold-refuse power over a frontier model behind a letter that was never made public, generalises: the ability to switch a capability on or off becomes a routine lever of sovereignty, exercised out of view. Decision systems are exempted from transparency law on security grounds. Procurement narrows to a few trusted vendors. The broad “capability-based” theory of control hardens, and with it the habit of deciding these questions in sealed rooms.
Citizens still experience the outcomes — routed, scored, served or not — but without visibility, without a route to challenge, and without standing to demand the record. This is synthocracy as sealed sovereignty: not the drama of machine takeover, but the mundane concentration of unreviewable decision power inside a stack that the public funds and cannot inspect. It can coexist with efficiency and even with a kind of order. What it cannot coexist with is contestable, accountable, democratic decision-making.
Signposts it is emerging: public-sector decision logic routinely exempted from freedom-of-information and audit on security grounds; export-control-style shutoff powers over deployed systems becoming normal instruments; agentic public infrastructure consolidating into a handful of vendors; sealed decisions defended on the ground that explaining them would itself be a risk.
Why this is foresight, and what it is for
None of these paths will arrive whole. Real states will run accountable agents in one department and ceremonial ones in the next, and seal a third under security law — the map is a patchwork, and the three pure types are instruments for reading it, not a prophecy about which one wins.
But the pure types share one revealing feature: the variable that most separates the accountable state from the other two is not capability, budget, or political colour. It is whether the record exists before the agent acts — whether identity, authority, delegation, and a real human boundary are in place ahead of the action, or reconstructed, if at all, after it. That is the admissibility question, and it is the reason our foresight loops back into our research rather than floating free of it. The scenarios describe where the road goes; the admissibility work is about the turn that is still ours to take.
The agentic state is not a distant prospect to be debated in the abstract. It is being built now, department by department, procurement by procurement. Which of these three it comes to resemble is not yet decided — and that, precisely, is the point of writing it down while it still isn’t.
FAQ
What is the “agentic state”? A state whose routine decisions — eligibility, licensing, inspection, prioritisation, flagging — are increasingly executed by AI agents that act, not merely by officials advised by software. The term marks the shift from AI that answers to AI that does.
Will AI run government decisions by 2030? By many industry projections, AI agents will handle a large share of routine government decision-making before 2030; adoption is already described as underway. What remains open is not whether but how — whether those decisions stay legible and contestable, drift into ceremonial oversight, or concentrate into sealed systems. Those are the three paths this foresight describes.
Is this a prediction? No. It is structured foresight: deliberately divergent scenarios built from observable trends, offered to clarify a choice, not to forecast an outcome. Real jurisdictions will most likely blend all three.
What decides which path a country takes? The single most decisive variable is whether the record — an agent’s identity, authority, delegation chain, and a real human boundary — exists before the agent acts. Where it does, power stays visible even as AI co-decides. Where it does not, decision power tends either to drift beyond reach or to concentrate behind a wall.
Sources (2026 baseline)
- Gartner, projections on government AI-agent adoption to 2028 and oversight requirements to 2029 (March 2026).
- Salesforce / IDC, Agentic Government survey (March 2026).
- IMDA (Singapore), Model AI Governance Framework for Agentic AI (January 2026; v1.5 May 2026); U.S. NIST Center for AI Standards and Innovation, agent-standards work (February 2026).
- EU AI Act Digital Omnibus — high-risk obligations deferred to 2 December 2027.
- Colorado SB 26-189 (automated decision-making technology): notice, adverse-outcome explanation, human-review rights.
- Case background: Australia’s Robodebt scheme; the Netherlands childcare-benefit (toeslagenaffaire) scandal.
- Synthocracy Institute companion papers: Admissibility for AI Agents: A Record-Based Test; The Eighteen Days: Admissibility at State Scale.
This is a Futures (foresight) publication of the Synthocracy Institute. It is clearly distinguished from our research strand, presents scenarios rather than findings, and is not offered as prediction or settled fact.
