In draft · STACK

The Agentic Codebase

Repositories built so AI agents become trustworthy collaborators

Demo an agent in an afternoon, regret it in six months. Build repos that last.

Prompts live in chat history, rules get copy-pasted across tools, and MCP servers sprawl without contracts, until the repo no longer survives a model upgrade or a second engineer. The Agentic Codebase introduces STACK, a five-layer framework for repositories that humans and agents maintain together. This is architecture, not a tutorial: the repo is the operating system, and every layer is versioned, tested, and owned like production code.

• The STACK framework: Structure, Toolchain, Agent configuration, Connection, Knowledge/quality
• Versioned patterns for AGENTS.md, CLAUDE.md, rules, skills, and subagent charters
• MCP server design, tool contracts, hooks, and guardrails that survive tool churn
• Agent memory tiers, context budgeting, evals, and CI for production discipline
• A copyable artifact per chapter: directory trees, SKILL.md, rule files, eval specs