How we build with coding agents ourselves

We don't just sell AI rollouts — we run our own development with agents. Not a pitch, but our actual day-to-day: how a feature goes from idea to merge request, which numbers changed — and where the honest limits are.

Three metrics — and half the truth

The loop behind every feature

Every feature takes the same path — here on a real ticket ("capture a screenshot in the chat"): from the idea to the finished merge request.

Implementing the example feature itself took around three minutes — including automatic translation into four languages and a finished changelog fragment. The work shifts from "writing code" to "giving context and checking results".

Retrieve · act · verify — in practice

Our setup is nothing other than the agent triangle in its purest form — each corner has its tools:

The third corner is the one almost everyone forgets — and the one we take most seriously: the agent starts a browser itself, clicks through the new feature like a user and checks end to end whether it really works. No human has to make the first click.

The stack behind it

• Live context: several MCP servers attached to Claude Code — Linear (tickets), Sentry (errors → fix), Context7 (current docs).
• One source of truth: a single AGENTS.md shared by every coding agent; CLAUDE.md only references it.
• Fast tools: moving from ESLint to Oxlint took the pre-commit check from ~1 minute to ~3 seconds — real leverage across many daily agent runs.
• Keeping it clean: Knip filters out unused code and exports so AI-written code does not become technical debt.
• Two minds: Codex for architecture, backend and security; Claude for frontend, UX and tool-calling — one agent conducts both and brings the results together.
• The last gatekeeper: CodeRabbit reviews every MR automatically, catching bugs and enforcing guardrails (style, permitted libraries, protected folders).

What this means for you

Coding was the first domain where agents went from adviser to doer — because it is deterministically checkable and reversible. We are currently transferring exactly that loop to the whole company: the agentic company. And where recurring work can be cast into agents and workflows, automation shows how.