Five composable layers that turn Claude Code from a code assistant into a governed, distributable agent runtime. Memory, knowledge, guardrails, delegation, and distribution — wired together and ready to ship.
The agent's persistent context. Loaded on every turn, never forgotten. Defines architecture rules, naming conventions, test expectations, and the repository map. Project scope overrides global scope.
Modular context packs invoked by description matching. The agent loads only what the current task needs — reference docs, scripts, templates — inside an isolated subagent so the main context stays clean.
Event-driven, non-AI policy enforcement — Git hooks for your agent. Events fire, matchers run, commands execute. No model in the loop, no ambiguity. The compliance layer for production agents.
Specialised workers with their own context, model, tools, and permissions. The main agent delegates and receives results — no infinite recursion, no context pollution. Code-reviewer, test-runner, explorer, anything you define.
Skills, agents, hooks, and commands packaged as one unit. Install across your team or publish to a marketplace. A package format for agent capabilities — versioned, distributable, reusable.
The bridge to the outside world. Any tool, any system, exposed to the agent through the Model Context Protocol — one open standard, no glue code per integration.
Multiple agents working in parallel with shared permissions and message passing. A lead orchestrates, specialists execute. The pattern that scales from prototype to production.
The full article and the PDF are member content. Magic-link login, no credit card, no risk — both available immediately.