Further explorations

Future directions

Speculative extensions and design explorations that build on the core platform. These are ideas under consideration, not current features.

The concepts on this page extend the Agentic Legibility Stack into areas that are not yet implemented. They represent directions for future development.
01

Publish once, serve everywhere

Departments author a single structured service description, and it is published automatically to every channel — web, API, phone script, AI agent, PDF, email.

The maintenance problem

Today, a single government service like Apply for Universal Credit exists as multiple parallel descriptions, each maintained by a different team: a GOV.UK web page, PDF and paper forms, phone script guidance, internal caseworker guidance, and API documentation. When eligibility rules change, five teams update five descriptions. They drift. Citizens get different answers depending on the channel.

Now add AI agents as a sixth channel, and the problem compounds.

5+
parallel descriptions per service
1,544
government services
~7,700
descriptions to keep in sync
1 → 5
one change = five updates
The question is not whether departments should support AI agents. It is whether they can afford to maintain yet another parallel description of each service — or whether this is the moment to describe services once and publish everywhere.

The proposition

Describe your service once, in a structured service description. From that single description, generate every channel automatically. Departments author structured truth. The platform generates channel-specific outputs. When something changes, departments update one place. Every channel reflects the change immediately.

Service description
Generator layer
MCP agent
Web form
API
Catalogue
Notify
Print
Departments don’t publish to channels. They describe their services. The platform publishes to channels.

Four dimensions of a service description

Each service is described across four dimensions. These are already defined in the current platform for AI agents — the publish-once model extends them to every channel.

Identity (manifest)

What the service is: name, department, eligibility summary, data it needs, SLA, complaint routes, escalation contacts. The service’s identity card.

manifest.json

Eligibility (policy)

Who is eligible and why: machine-readable rules with human-readable failure reasons. “You must be over 16”, “You must have a valid driving licence”. Edge cases documented.

policy.json

Journey (state model)

What the journey looks like: a step-by-step map from “not started” to “completed”. Which steps are required, which are optional, what unlocks what. The service’s choreography.

state-model.json

Data sharing (consent)

What data is shared and why: every piece of data the service needs, who it comes from, what it’s used for, how long it’s kept, and how to revoke. The service’s data contract.

consent.json

Output targets

From the same service descriptions, the platform generates channel-specific outputs for every delivery surface.

Channel What it generates Status
AI agent (MCP) Tools, resources, prompts for Claude/GPT Built
Web form (GOV.UK Forms) Pages, questions, routing, confirmation Feasible
API (OpenAPI 3.0) REST endpoints, request/response schemas Straightforward
Service catalogue Schema.org/GovernmentService + DCAT entry Trivial
Notifications (GOV.UK Notify) Email/SMS templates per journey stage Feasible
Accessible summary (HTML/print) Printable service overview page Feasible

How this aligns with GDS standards

Publish once is not a departure from existing government digital standards. It is a direct implementation of them.

Point 3: Joined-up channels

One source of truth means citizens get the same answer whether they use the web form, the AI agent, or the phone script generated from the same description.

Point 13: Open standards

Service descriptions are JSON, output includes OpenAPI 3.0, Schema.org, DCAT. Every format is machine-readable and standards-based.

Technology Code of Practice

The generator layer is reusable across all departments. Build once, use 1,544 times.

API Technical Standards

OpenAPI 3.0 is the government-mandated format. The platform generates compliant specs automatically from service descriptions.

International context

Three countries have demonstrated, at national scale, that government services are data — not websites.

Estonia — X-Road

929 institutions, 3,000+ digital services, two decades of secure decentralised data exchange. X-Road proved that services can be described as interoperable, machine-readable specifications and federated across organisations.

Singapore — SGTS

Composable modules — identity, payments, notifications — assembled into services. MyInfo’s deployment time dropped from one year to four months. “Describe once, compose everywhere” at national scale.

Ukraine — Diia

23 million users, 130+ services. Diia.AI launched September 2025 as the world’s first government AI agent delivering real services via MCP. Ukraine coined the term “Agentic State”: citizens express intent, AI executes.

Three countries, three architectures, one convergent insight: government services are data, not websites. Describe them as structured specifications, and every delivery channel — including AI agents — becomes a rendering target.

The interaction typology

13 standard service shapes have been identified across 1,544 government services. Each shape comes with a pre-built state model template — the skeleton of the journey. Departments select the closest shape, then customise.

Type What it means Example services
Benefit Citizen seeks financial support; income-assessed Universal Credit, PIP, Housing Benefit
Application Submit information, receive a decision Passport, Child Benefit, Blue Badge
Licence Prove permission to do something Driving licence, fishing licence
Obligation Something the citizen must do Update DVLA address, notify OPG of death
Legal process Navigate complexity with professional support Probate, LPA, divorce
Portal Access multiple services through one account UC journal, student loans

Plus 7 more shapes covering registrations, entitlements, grants, payments, appointments, task lists, and information hubs. Departments don’t design service journeys from scratch — they select a shape, fill in the specifics, and the platform generates the journey for every channel.

What this makes possible

Describing services as data is not the end state. It is the precondition for three capabilities that are currently impossible.

Proactive services — if a citizen’s circumstances change (recorded through one service), the platform can identify other services they’re now eligible for and notify them automatically.

Life event orchestration — 16 life events have been mapped across all 1,544 services. When a citizen experiences a birth, bereavement, job loss, or retirement, the platform identifies all relevant services and sequences them.

Departmental analytics — when services are described as data, departments can measure coverage, identify gaps, compare interaction patterns, and understand where citizens get stuck.

Publishing once is not a technical convenience. It is the precondition for joined-up government. When services are data, joining them up is computation.
This section describes a target operating model. The AI agent channel is built and operational — the remaining channels require generators, not new data. See the platform overview for what is implemented today.
02

Interaction model explorations

Four distinct UI paradigms were explored for the citizen app. Each addresses a fundamental question: what should your relationship with government look like on a phone?

The current model

The current app uses a traditional four-tab structure: Home, Chat, To-do, and Wallet. Despite having transformational agentic capabilities — life event detection, cross-service field merging, delegated action, proactive eligibility — the interaction model still organises itself around government’s taxonomy. Services live in one tab, conversations in another, tasks in a third. The agentic loop is fractured across views.

Four alternative models

The Mirror Recommended

You open the app and see your life — your family, your money, your health, your vehicles. Government is invisible infrastructure. The app mirrors your life back to you, organised into facets. Data-driven alerts surface what needs attention. Conversations open from data points and live inside your life facets, with the depth and persistence of full threads. Life event journeys unfold as swipeable stories.

The Journal

Conversations are the entire app. Your relationship with government is a series of threads — some brief, some spanning months. Plans, tasks, credentials, and consent all live inside conversations, not in separate tabs. There is no service directory because you never browse; you just talk. The agent initiates threads proactively when something needs attention.

The Stream

A living, intelligent feed. No pages to navigate to — just a single stream of cards representing everything happening in your relationship with government. Active journeys appear as stories at the top. Priority items are pinned. Everything else flows through a timeline. Every card is actionable, and any card can expand into a conversation.

Comparing the models

The Journal

Best when citizens have complex, ongoing needs. The conversation is the relationship.

MetaphorMessaging app
DiscoveryAgent starts threads
ConversationsFull-depth threads
StrengthContinuity and context
Best forPower users, complex journeys

The Stream

Best when immediacy matters. Everything is one scroll away. No navigation decisions.

MetaphorIntelligent feed
DiscoveryEligibility cards in feed
ConversationsExpandable from cards
StrengthImmediacy, zero navigation
Best forCasual users, quick actions

Current (tab-based)

Traditional four-tab structure. Familiar but fractures the agentic loop across separate views.

MetaphorService directory
DiscoveryBrowse categories
ConversationsSingle, in separate tab
StrengthFamiliarity
Best forKnown service lookups

Why the Mirror was recommended

The recommended model is a hybrid that takes the strongest idea from each: identity-first organisation from the Mirror, conversation depth from the Journal, story progression for life events from the Stream. The organising principle is simple: your life comes first, government wraps around it.

All four models share one principle: the tab bar is gone. Government’s taxonomy is invisible. Conversations are contextual, not isolated. The agentic loop — from life event detection through to credential issuance — flows naturally rather than being fractured across separate views.
These interaction models are design explorations, not implementation commitments. The current citizen app uses the tab-based model. A future iteration may adopt elements of the Mirror hybrid.