AI Coding Agents

AI coding agents are tools that read a repository, edit files, run commands and tests, and propose changes the way a developer would. Our engineers use four of them every day: Claude Code, Codex, Pi and OpenCode. Agents take the routine part of the work (tests, refactors, migrations, boilerplate, documentation), and engineers keep what decides a project’s outcome: architecture, review of every change and responsibility for what ships.

What Agents Do in Our Projects

Agents work on tasks with a clear definition of done. They write unit and integration tests for existing code, which is often the fastest way to make a legacy module safe to change. They carry out mechanical refactors across hundreds of files: renaming a domain concept, upgrading a framework version, replacing a deprecated API. They draft database migrations, API clients from an OpenAPI spec, admin screens and the glue code between services. They read unfamiliar codebases and produce a map of modules, data flows and risky spots before we estimate a change. On the same footing they help with the tedious parts of Rust and Tauri projects, where the compiler catches most mistakes in generated code before a human even looks at it.

We do not hand agents product decisions, security-sensitive design or anything that cannot be verified by tests and review. The goal is more engineering time spent on the problems that need judgment.

How We Work With Agents

Every repository where we use agents gets an instruction file (AGENTS.md or CLAUDE.md) that describes the stack, commands, conventions and what must never change. Agents read it at the start of each session, so they follow the project’s rules instead of guessing.

Work starts from a task in the tracker. For anything larger than a small fix, the agent first proposes a plan: which files change, in what order, how the result will be checked. An engineer approves or corrects the plan before any code is written. The agent then makes the change, runs the build, linters and tests, and the engineer reviews the diff line by line, exactly as they would review a colleague’s pull request.

For important changes we use a second agent from a different vendor as an independent reviewer. A diff written with Claude Code can be checked by Codex, and the other way round. Different models make different mistakes, so cross-review catches problems a single model tends to repeat. Its findings go to the engineer, who decides what to fix.

The final gate is the same as for human-written code: CI must pass, and nothing reaches the main branch without an approved review.

Your Code and Data

Before agents touch a client repository, we agree on the rules. Which agents and model providers are allowed, whether code may leave your infrastructure, which directories are off-limits. Secrets, production credentials and personal data are never put into prompts. When a client requires a specific provider, an enterprise agreement or a self-hosted model, we use OpenCode or another agent configured for that provider. If agents are not allowed on your project at all, we work without them.

When This Approach Fits

Agent-assisted development pays off most on projects with a lot of well-defined work: test coverage for an existing system, framework upgrades, migrations, internal tools and integrations, and new features in a codebase with clear conventions. It gives less on greenfield architecture work and on research-heavy tasks, where the hard part is deciding what to build. For AI features inside your product, such as assistants, search over documents or process automation, see our AI development and LLM integration pages.