Sovereign AI-Native Engineering

The choices we made, and why they matter in a regulated estate.

KR&A’s methodology puts AI agents at the centre of software delivery. We run it on a stack chosen against one rule: nothing the client depends on can be withdrawn, repriced or reached by a foreign court through a decision the client did not make.

The exposure

Three ways a toolchain fails a regulated buyer without anyone touching it.

Licence change

The terms move after adoption.

In 2021 Docker Desktop became paid for organisations above 250 employees or US$10 million revenue. In 2023 Terraform moved to the Business Source Licence. Neither needed the customer’s consent.

Jurisdiction

Data residency is not data control.

A provider subject to US law can be compelled under the CLOUD Act to produce data it controls, wherever it is stored. What matters is who can be compelled, not where the server sits.

Model drift

The vendor holds the compliance posture.

A proprietary model can change without notice, and each silent update is a de facto re-certification event. Open weights pinned to a version change only when the client decides.

The test

Two limbs. Both must hold.

Every component was selected against the test KR&A applies in client sovereignty assessments. Domicile alone is insufficient: it records where a company is registered, not where the data goes or who can withdraw the capability.

Limb 1 · Forkability

Forkable under an OSI-approved licence.

The capability must be runnable indefinitely without the vendor. Open core fails where the feature that matters sits behind the commercial gate.

Limb 2 · Runtime control

Jurisdictional control of data flow at runtime.

At execution time, code, prompts, credentials and telemetry traverse only infrastructure in a jurisdiction the client selects. A vendor control plane inside the client’s account fails this limb.

Every choice, and why

One stack, each component replaceable, none of it rented from outside the EU.

LayerChoiceLicenceWhy
Agent harnesspiMITKR&A’s skills and lifecycle rules are files committed with the code. Any compatible model endpoint plugs in.
Agent terminalTUIOSMITAgents run side by side in persistent sessions, and every request for human approval lands in one inbox.
ModelsMistral, GLM-5.2, QwenApache 2.0, MITOpen weights pinned by version. Movable to self-hosted inference at any time.
InferenceScaleway Generative APIsServiceHosted in Paris. Dedicated deployments where single-region processing is required.
CloudScalewayServiceFrench provider. Compute, PostgreSQL, storage and edge under one EU contract.
EdgeScaleway Edge, DNSServiceRemoves the last US operator from the request path.
Source, CIGitLab, self-hostedMIT (CE)Forge, pipelines and registry, run on our own infrastructure inside the estate.
InfrastructureOpenTofuMPL-2.0Linux Foundation fork of Terraform. Same language, open governance.
Orchestrationk3sApache 2.0Lightweight CNCF Kubernetes that runs the same on any provider.
DeliveryFlux CDApache 2.0Every production change is a reviewed commit the cluster pulls.
SecretsOpenBao, ESOMPL-2.0, Apache 2.0One store for platform and developer secrets, injected at runtime.
Work trackingPlane, self-hostedAGPL-3.0System of record for scope, agent activity and decisions. See governed delivery
ContainersColima, Docker EngineMIT, Apache 2.0Drops the proprietary desktop layer, keeps the standard Docker toolchain.
ObservabilityGrafana, Loki, VictoriaMetricsAGPL-3.0, Apache 2.0Logs and metrics stay inside the estate they describe.
Workstationchezmoi dotfilesMITOne declared inventory installs every tool and the manifest agents read.
On model origin

GLM and Qwen are developed in China. Under the test, origin is context rather than verdict. The weights are published under MIT and Apache 2.0, run on infrastructure in Paris operated by a French provider, and send nothing to their developers. Where a client’s policy prefers European-origin models, Mistral takes the primary role and the harness is unchanged, because the model is a configuration value, not an architectural commitment.

Governed delivery

From the outside, an agile project. Underneath, a complete provenance record.

Non-technical stakeholders see what they already know: a backlog, iterations, story points and a board in Plane. They need no new tooling and no new vocabulary to follow progress.

Behind each card, the work item holds the specification, the agent’s plan, every deviation, the verification evidence and the human decision that closed it. The harness enforces this lifecycle through committed rules, not individual discipline.

Work-item lifecycleEnforced by the harness
  1. 01

    Filed

    The work item enters the current iteration with an estimate and acceptance criteria. Filing is not authorisation to build.

  2. 02

    Planned

    The agent moves the item to In Progress and posts its implementation plan before changing any code. Reviewers diff the result against that plan.

  3. 03

    Logged

    Deviations, root causes and discovered constraints are recorded on the item as they occur.

  4. 04

    Evidenced

    At In Review the item carries the change reference, pipeline result and verification evidence, including visual confirmation for anything a user sees.

  5. 05

    Closed by a human

    No agent moves its own work to Done, even when every check passes.

  1. Work item
  2. Branch
  3. Merge
  4. Pipeline
  5. Deploy
  6. Live
For regulated environments

The evidence comes out of delivery, not a separate exercise.

ObligationWhat the process produces
DORA, ICT third-party riskEvery component is forkable with a documented substitute, which gives exit strategies substance rather than intent.
NIS2, supply-chain securityA declared toolchain inventory, digest-pinned images and a single path from commit to production.
Cyber Resilience ActTraceability from requirement to released change, recorded as a by-product of normal delivery.
EU AI Act, Article 4A defined operating model in which named people supervise agents and own every decision to ship.

These map the evidence the process generates. They are not legal advice and not a compliance opinion. Confirm obligations with counsel.

Status

We run on it first.

KR&A’s own platform is the reference implementation. It is migrating from a US hyperscaler stack to the components above, and the migration is itself run through this process, with every step recorded in Plane.

Engagement

Assessment first.

We baseline your development toolchain against the sovereignty test and give you time to absorb the result. It converts into a scoped pilot on one workstream, then training and adoption, with no upfront commitment beyond the assessment.

See it in action

KR&A’s own platform is built this way: seven sites, one monorepo, every change taken through the work-item lifecycle above and recorded in Plane. The architecture decisions and infrastructure guides are published in full.

docs.kevinryan.io ↗