Components.

A tooling map for state-grade agentic systems, grouped by layer: what has been observed in referenced public-sector PoCs, what to add, and what to weigh as an alternative. Names link to the projects. None of this is an institutional endorsement, and none of it changes the rule that the public service decides.

In use: Observed in referenced public-sector PoCs
Add: Recommended addition for state-grade agentic use
Evaluate: Alternative to evaluate

Agent orchestration, harnesses, protocols

In use: LangGraph
In use: LangChaincreate_agent, middleware
Add: deepagentsplanning, filesystem context, subagents, skills; Python and JavaScript
In use: DSPy
In use: MCP specification
In use: FastMCP
In use: langchain-mcp-adapters
Add: MCP Registrybuilding-block discoverability
In use: A2A protocol
In use: a2a-sdkPython
Add: Agent Skills specificationportable skill folders, used by deepagents
Add: Claude Agent SDK
Evaluate: Pydantic AItyped agents; pairs with pydantic-evals below
Evaluate: Google ADK
Evaluate: OpenAI Agents SDK

LLM access, inference, embeddings

In use: LiteLLM proxy
In use: vLLM
In use: OpenRouter
In use: Azure AI FoundryEU regions
In use: FastEmbed
In use: multilingual-e5-large
In use: BGE-M3
Add: Ollamalocal development

Retrieval, storage, document parsing

In use: Qdrant
Evaluate: pgvectorif collapsing to one datastore
In use: SQLAlchemy
In use: Alembic
In use: pypdf
In use: python-docx
Add: Doclinglayout-aware PDF and DOCX to structured output; replaces ad-hoc parsing
Add: OCRmyPDF with Tesseract
In use: WeasyPrint
In use: PlaywrightHTML to PDF, end-to-end tests
In use: PptxGenJS
In use: SearXNG

PII, guardrails, Estonian NLP

In use: Microsoft Presidio
In use: EstBERTTartuNLP named-entity recognition
In use: EstNLTK
Add: NeMo GuardrailsColang rails on gateway egress
Evaluate: LLM Guardprompt-injection and leak scanners as LiteLLM middleware

Speech and voice

In use: Speechmaticsbatch and real-time speech to text with EU processing; English and Estonian; the loop on /try transcribes with it · example: medical transcription

Evaluation, tracing, observability

In use: MLflowtraces, evaluations
In use: pydantic-evals
Add: Langfuseself-hosted, OpenTelemetry-native; data stays in a national cloud such as Riigipilv, where LangSmith is SaaS in the US
In use: OpenTelemetry GenAI semantic conventions
In use: Prometheus
In use: Grafana Loki
In use: Grafana Alloy
Add: Inspect AIthe UK AI Security Institute’s evaluation framework
Add: Promptfooregression and red-team checks in CI
Evaluate: RAGASretrieval-specific metrics

Identity, authorization, secrets, signing

In use: OpenFGA
In use: SPIFFE/SPIRE
In use: HashiCorp VaultTransit
Evaluate: OpenBaoMPL fork of Vault; procurement-friendlier licence
In use: TARA
In use: Authlib
In use: Web eID
In use: undersignASiC-E and XAdES in Node
Evaluate: digidoc4jRIA-maintained, Java; if a JVM signing service is acceptable
Add: EUDI Wallet reference implementationthe citizen-agent SD-JWT VC path
Add: sd-jwt-python
Add: OpenID Shared Signals and CAEPpush revocation
Add: OWASP Top 10 for LLM Applications

Interoperability, GovStack, registries

In use: X-Road (NIIS)
Add: GovStack specifications
In use: RHR, the public procurement register
In use: RHR open data
In use: TED API
In use: Riigi Teataja
In use: Äriregister open data

Jobs, runtime, API

In use: FastAPI
In use: sse-starlette
In use: Pydantic v2
In use: ProcrastinatePostgres-backed queue
Evaluate: arqRedis
Evaluate: Temporaldurable execution for multi-day agent runs; evaluate only if LangGraph checkpointing proves insufficient
In use: httpx

Frontend

In use: Next.js
In use: TanStack Query
In use: Tiptap
In use: Radix UI
In use: Tailwind
In use: Recharts
In use: Astro
In use: axe-coreWCAG checks in CI

Development tooling

In use: uv
In use: ruff
In use: pytest
In use: Vitest
In use: import-linterenforces the dependency matrix
In use: Docker Compose
In use: Ansible
In use: Terraform

Agentic Government Reference Architecture