9. Tech Stack
Leapfrogging to tomorrow’s enterprise stack.
How It (Doesn’t) Work Today
Over the past two decades, the ideal government technology stack has emerged: Its building blocks include standardised modules for identity, data exchange, messaging, user interface, payments, and platforms for operating and designing registers and services. All of those blocks run cloud-natively on modern infrastructure.
This stack allows for economies of scale while abstracting technical complexity away from most public bodies, enabling them to develop functional services at a lower cost and with greater speed (consider the India Stack, for example, X-road in Estonia or Ukraine’s platform of registers). Increasingly, these building blocks are also being standardised and developed internationally as digital public goods (think of the Nordic Institute for Interoperability Solutions or the GovStack initiative, for example).
The truth is less flattering. Most governments are a decade away from achieving enterprise best practice in the pre-Agentic AI era Some of the gaps are as follows:
Infrastructure and operational gap: Workloads are still hosted on fixed virtual machines (or worse, 1990s servers). Systems are sized for peak usage. Common enterprise IT operations best practices are, such as DevOps rather the exception, whereas private sector DevOps penetration is over 80 percent. Budgeting is typically capital-expenditure heavy and oriented toward maintaining legacy infrastructure rather than developing modern, user-focused applications.
Architecture and shared platform gap: Applications often remain monolithic and agency-specific Integration happens through brittle, point-to-point calls. Few countries have adopted data-sharing platforms like Estonia’s X-Road, which enables secure cross-agency data exchange. APIs for essential functions, such as ID, payments, and secure messaging, are rare, meaning cross-agency workflows require custom workarounds.
Openness gap: Open-source codebases for government software, vendor-neutral standards, and discoverable open data (and open APIs for restricted data) remain the exception rather than the norm.
Policy gap: Laws, policies, and entitlements are generally published as prose rather than in machine-readable formats. Without codified rules, automation, transparency, and trust become harder to scale. Policy registries are not designed to be auditable or consumable by AI agents, limiting the ability to build intelligent, rule-based public services.
Together, these gaps lead to the massive cost overruns and dramatic failures common to nearly all countries’ public sector technology infrastructure.
A Vision for an Agentic Government Tech Stack
Even as most administrations struggle with foundational IT, the enterprise (and by extension, government) tech stack will need to be redesigned from the ground up to take advantage of the capabilities of agentic AI.
This will entail a rethink (and most likely a re-engineering) of nearly every existing functional layer of the government tech stack:
Schematic view of an agentic government tech stack
Building blocks for an agentic government tech stack
Building Block | What It Means in Practice | Why Agentic Agents Need It |
Compute infrastructure | Serverless and agnostic. A model router picks the best model per task and records cost and energy usage per call. | Keeps inference fast and affordable, and prevents public agencies from buying or renting half-idle hardware. |
Simulation or digital-twin sandbox | A live mirror of critical systems where policy tweaks or code updates run safely before hitting production. | Allows for testing rule changes, loading spikes, or cyber-attacks without real-world fallout. Provides evidence before action. |
Identity 2.0 | A dual identity system: (i) Human: verifiable credentials for people, businesses, and officials. (ii) Machine: cryptographically signed IDs for software agents, delegation tokens spelling out who may act for whom, on which tasks, for how long. | Every agent transaction starts with the question “who am I, on whose behalf, and am I authorised?”. Clear, revocable credentials and delegation enable humans to stay in control and allow auditors to trace actions. |
Programmable payments rail | Tokenised infrastructure that enables the public sector to support smart contracts, such as escrow and rules-driven payouts, as well as facilitating the straightforward integration of payment APIs. | Lets agents make and receive payments, with no manual finance step required. E.g. pay EUR 0.38/km for verified snow-ploughing routes, or trigger subsidies on verified milestones. |
National data commons | A combination of a large-scale data lake and a data mesh, where all data (from high-value registers to operational logs) are catalogued. There is a retrieval layer for quick calls. Functions to generate for anonymisation and generating synthetic data. Fine-grained data-governance policies track consent, purpose and lineage for every query, while also enabling privacy-preserving machine learning to operate across the entire government data set. Laws, policies and KPIs exist in machine-readable decision tables allowing dynamic rewriting. Each version is tested, signed, and published alongside the legal prose. | Agents learn patterns and answer questions only when they can safely access real-world data. Lineage proves nothing was misused. |
Service orchestration and model context protocol (MCP) | Cross-government fabric to orchestrate the interaction between different public and private sector agents, with a machine- readable directory of every API and capability. A workflow engine enables agents to combine tasks in new ways. | Enables ‘Lego-brick’ composability, allowing agents to rapidly build new life-event flows by snapping tasks together. |
User interface | One multimodal shell (chat, voice, web, AR, visual) for accessing the orchestration fabric. It offers accessibility features (speech-to-text, multilingual) as well as an API for agents to appear embedded inside third-party apps. | Government meets users where they already are, whether that is with smart glasses or ERP systems. No more hunting for portals. |
There will also be new functional layers, notably for governance:
AI evaluation and transparency layer: Provides user-visible real-time logs, model ‘nutrition labels’, software bills of materials, and public dashboards showing uptime, decision volumes, and appeal rates. Converts the black box into a glass box, by allowing citizens and auditors to see, question, and correct agent behaviour.
Agent registry and governance: Acts as a control tower for all authorised agents, recording their scope, audit hooks, rollback history, and kill-switches. Includes sandbox, certification, and marketplace functions for third-party agents.
To guide architectural decisions, one can think of the tech stack as comprising three concentric systems, each with a distinct purpose and optimal ownership model:
Compute substrate: This is the foundational horsepower layer: cloud infrastructure, GPU clusters, and scalable data storage. Its role is to provide elastic capacity to train, fine-tune, and run AI agents at speed and scale. The operating model here is market-driven: governments should rent what already exists, prioritising portability and strong SLAs.
Agentic Digital Public Infrastructure: These are the reusable building blocks every agent and service depends on: identity, secure messaging, payment rails, data catalogues, and task-level APIs. These blocks should be co-developed via public–private joint ventures or open-source consortia. Think of an open-source ID wallet, or a cross-border API standard for public service agents.
Sovereign governance layer: This is the crown jewel, the part of the system that encodes law, accountability, and democratic control. It includes rules-as-code, agent registries, audit mechanisms, redress procedures, and kill-switches. Because this layer defines what is legally binding, what agents may act, and how decisions can be contested or rolled back, it must remain under public ownership and stewardship, even if some tools or frameworks are shared across countries or sourced from the market.
Key Questions
How can governments commit to long-term architectures when technology standards, tooling, and practices are evolving rapidly? What governance principles or ‘minimum bets’ ensure flexibility without paralysis?
What straightforward, high-level checks can help CTOs identify infrastructure investments that are incompatible with agentic AI? How should governments audit or phase out platforms that cannot provide task-level, agent-ready interfaces within a defined timeframe? What policy mandates and infrastructure shifts are needed to ensure AI workloads run in elastic, certified cloud environments?
What governance and procurement strategies should guide the use of commercial cloud and compute providers, given that compute is a market-based layer? How can governments ensure portability, SLAs, and data protection when renting this foundational infrastructure?
Which public–private collaboration models are best suited for building shared agentic digital public infrastructure, such as identity systems or task-level APIs? How can governments coordinate across sectors and borders for the sake of interoperability?
What safeguards and design principles must apply to the sovereign governance layer, where law, audit, and redress are encoded? How can governments retain full control over this layer while reusing open tooling or cross-national standards?