How we build software

AR Labs builds most of its software with AI coding agents: AI agents that write code, such as Claude Code or Codex. A person decides what to build and approves the plan; agents write the specifications, the tests and the code; a second model reviews the result in a fresh context; and a person merges it. The pages in this section describe that method one practice at a time, for an engineer or an AI agent who wants to adopt some of it.

The method is spec-driven and test-first. Every piece of work is tracked as a node in a dependency graph of tasks, gets its own written specification, is built on its own branch in an isolated checkout, and must prove each acceptance criterion with a committed test, a committed check script anyone can re-run, or a screenshot a person can judge. A claim that the work is done counts for nothing on its own.

Each rule here exists because something went wrong without it. A review step that lived only in written instructions was skipped by an unattended run. A warning saved in one checkout was invisible from every other. A fix closed one race condition and left three identical ones in a neighbouring class. Each page names the problem a rule prevents alongside the rule, because the problem is what lets you judge whether the rule fits your own work.

When a written rule kept being broken, we moved it into a script: a hook, which runs automatically around an agent's action and can block it, or a check the work must pass before it moves on. The written rule stays, to explain why the script exists.

If you are new to the section, start with how AR Labs engineers software with AI agents, the approach on one page. Then read the coding pipeline, which shows how the practices fit together step by step. The overview pages after it cover agentic engineering, architecture, unattended runs, human-in-the-loop gates and multi-agent orchestration; the practice pages each take one rule in detail, from spec-driven development to evidence of completion and worktree isolation.