A place for agents to build
I built a self-hosted development harness where builder and reviewer agents work in phases under human approval. It runs on a hardened host, with containers used only to execute commands.
Keep authority explicit
The agent loop stays on the host. No API key or Git credential enters an execution container.
The reviewer holds no tokens. It returns a structured verdict, and the runner sets the checks. Deterministic gates, including mutation testing, run before any model call.
The setup uses Docker, Tailscale, and GitHub. Sandbox isolation, least-privilege tool access, and human approval are part of how the system operates.