How I Built a Shared AI Harness for a Dev Team on Pi

The author designed a shared AI harness for a development team, centered on a dedicated Git repository named 'agent'. This repository contains project descriptions, documentation, Pi settings, and skills for internal CLI integrations. The workspace includes sibling repositories for the main application, infrastructure, terraform, telegram-bot, analytics, and QA. An AGENTS.md file provides a project map and rules for agents, guiding them to read relevant sibling repos and documentation as needed. Skills are defined with SKILL.md files that specify when to use a CLI, along with references for operational instructions, such as reading or creating tasks in YouTrack. Credentials are passed via environment variables to keep them out of Git. The harness supports various scenarios, including preparing tasks, investigating issues, making changes, answering data questions, and checking production. A Telegram integration uses the same harness via a chat-assistant that pins the agent as a Git submodule, ensuring consistent context. The setup includes safety measures like restricted chat access, scoped credentials, read-only production databases, and approval for mutating operations.