A government chatbot can retrieve a form or summarise a regulation. An autonomous system can assemble evidence, choose a sequence of tools, initiate a workflow and return with a completed result. That additional verb, act, changes the institutional question entirely.
Across 2026, governments have begun to describe their AI programmes in the language of agents. The label is applied generously. Microsoft Copilot licences, research assistants, procurement notices and genuinely self-directed systems often appear in the same category. A useful comparison therefore begins with evidence: the authority delegated to the system, the tools it can reach, the consequences of an error and the point at which a person must intervene.
Autonomous adoption is still early. The most credible national programmes remain bounded by sandboxes, approval gates and narrowly defined tasks. Yet the direction is visible. Public administration is moving from software that produces language towards software that participates in operations.
Singapore: autonomy begins with permission
Singapore has taken the clearest governance-first route. Its Cyber Security Agency, GovTech and Infocomm Media Development Authority opened an AI Agents Sandbox with Google in August 2025 and published the findings in May 2026. The tests covered computer-use agents in quality assurance, AI safety and social assistance. They also exposed the practical problem beneath the theory: identity, authentication and permissions designed for human users become inadequate when a machine can operate a browser or call several systems on a person's behalf.
That experience shaped Singapore's Model AI Governance Framework for Agentic AI, released in January 2026. The framework asks organisations to limit the agent's tools and data, define approval checkpoints, whitelist actions, test predictable failure modes and disclose what the system can do. In May, the government stated that public-facing agents should not perform high-stakes or irreversible actions without human review.
Singapore's achievement is institutional clarity. It has treated autonomy as an access-control problem as much as a model problem. The programme remains controlled rather than ubiquitous, though its sandbox provides unusually concrete evidence of systems being tested against real workflows.
China: standards before legibility
China is building the connective tissue for scale. In May, the Cyberspace Administration described agents as systems capable of autonomous perception, memory, decision, interaction and execution, then mapped nineteen application areas spanning government approvals, judicial support, city management, procurement, healthcare and tourism. In July, the national standards authority announced seven standards covering architecture, identity, description, discovery, interaction and tool invocation.
The standards effort involves more than seventy organisations and an industry initiative with more than one hundred companies. Official accounts refer to over fifty pilot applications and testing in Beijing's Haidian district. The ambition is larger than a collection of departmental experiments: agents should be discoverable, interoperable and attributable across a national digital environment.
The evidential weakness is equally important. Aggregate pilot counts reveal little about delegated authority, measured outcomes or production use. China currently presents the strongest machinery for standardisation and one of the least transparent views of operational performance. Its model may scale quickly once common identity and invocation rules settle, but the public record still makes maturity difficult to audit.
Abu Dhabi: the executive layer arrives first
Abu Dhabi offers the most direct senior-government example. On 17 September, the Executive Council used an AI Committee System, described by the government as an agentic platform, to support analysis and data-led discussion of public projects. The announcement does not show the platform making binding decisions; it does place an agentic system inside the room where decisions are formed.
A second programme is more operational. The Hazardous Materials Management Centre and Presight are developing agents that connect incident reports and operational indicators, detect anomalous patterns, assess priorities and recommend actions across the lifecycle of hazardous materials. Prototypes will be measured against detection speed, alert accuracy, response time and automation levels, with human oversight and data sovereignty built into the design.
Abu Dhabi's advantage is concentrated execution: government, data infrastructure and major suppliers can move in concert. Its risk is definitional inflation. A rollout of Microsoft 365 Copilot to 35,000 employees is important digital capacity, though it should not be confused with autonomous administration. The council and hazardous-materials cases are more valuable because they specify where agency enters the workflow.
United States: autonomy through the laboratory
The American federal route is mission-led. The Genesis Mission is creating agents that test scientific hypotheses and automate research workflows across federal datasets. A July funding package of more than $5 billion extended the plan to agentic analysis of 150 petabytes of NASA and Department of Energy data, agentic engineering workflows and autonomous laboratories where robotics, edge AI and real-time analysis form closed experimental loops.
This is substantial autonomy, although it sits in science rather than citizen services. A laboratory agent can propose an experiment, coordinate instruments, interpret results and select the next run within technical and safety limits. Its outputs remain accountable to scientists, but the machine receives genuine latitude over sequence and execution.
Elsewhere, the federal government is still assembling procurement, training and assurance. The General Services Administration is training employees in agent architecture and orchestration while expanding access to frontier models. National-security policy stresses controllability and a clear chain of command. The American pattern is therefore asymmetric: advanced autonomy in research and security, broad assistants across administration, and limited visible delegation in public-facing services.
India: a named agent inside a sovereign platform
India provides one of the most specific procurement cases. In May, the Comptroller and Auditor General sought a sovereign AI and data platform with a dedicated pillar for agentic applications. The tender names AI-PARAS, a pension pre-scrutiny agent intended to examine cases before approval.
The distinction between a tender and a deployment matters. AI-PARAS is evidence of institutional intent, budget architecture and a defined workflow, rather than proof of a mature live service. Even so, the design is revealing. India is attaching an agent to a repetitive, rules-heavy administrative process where documents are numerous, errors are costly and final authority can remain with a human officer.
This may become the practical emerging-market template: sovereign infrastructure, modular agents and clearly delimited administrative jobs. It demands less theatrical capability than a universal government assistant and offers a cleaner route to measuring time saved, inconsistencies found and cases escalated.