Agentic Engineering
AI agents support us across engineering, from exploring approaches and implementing changes to testing, review, and documentation.
We apply them with software-engineering expertise, safely and with genuine digital sovereignty: security, privacy, and control come before convenience.
For clients, that means faster feedback and more engineering capacity without giving up quality, security, or control.
How we work with agents
Humans own the outcome
An agent can write a patch, but it cannot take responsibility for one. We review every change and stay accountable for what reaches production, so there is always a named engineer behind it.
Bounded autonomy
Agents work on scoped tasks inside the repository, on a branch, in an isolated environment with the least privilege they need. Nothing reaches production without a human merging it.
Verified, not trusted
Output is judged by automated checks: tests, type checks, linters, reproducible builds. If a change cannot be verified mechanically, it does not go out.
Sovereignty over convenience
Agentic tooling should not turn speed into dependency. We keep workflows portable and data protection an engineering constraint. See digital sovereignty and the providers behind it in our Trust Center.
Secure by default
Agent work runs sandboxed and isolated, with no production access or credentials in reach. Client personal data stays out of our internal AI tooling, and every provider we use is listed in our Trust Center.
Reproducible tooling
Pinned toolchains and Nix-based environments mean an agent's checks run against the same dependencies every time, so results are comparable instead of anecdotal.
Where this shows up
- In our own tooling. Our open-source opencode-decisions plugin adds screening, secret scrubbing, and decision tools to the agent tooling we use ourselves. It is a concrete example of how we build guardrails around agentic work.
- In client systems. When a client wants to adopt agentic workflows, we start with the guardrails and the verification, not with the prompts.
- In the same standards as everything else. Typed code, reproducible builds, and a CI pipeline that gets the final word. See our values.