BotBytes OS - Multi-Agent AI Operating System for an Agency

Claude Code is Anthropic's coding agent. Out of the box it is a very capable generalist. The work is making it follow your rules reliably, every session, without being reminded.

My own agency operating system, in daily use: a portable instruction system that routes rather than instructs. The root file only points — it holds no task detail. Underneath it sit rule files for each class of decision, skills whose bodies load only when invoked (so they cost nothing until used), subagents for parallel work, and a three-tier memory index so the right note is found instead of everything being loaded.

Hard gates implemented as hooks, not instructions. Some actions are blocked by code before they run — a command touching a protected path is refused, and a budget check stops work once a ceiling is crossed. An instruction can be forgotten; a hook cannot.

Approval gates. Anything irreversible stops and waits for a human, by design.

Every rule carries a date and the reason it exists, so the system explains itself to the next person who reads it.

Tool-agnostic. The same instruction tree is mirrored for other harnesses, so the work is not locked to one vendor.

Claude Code, Claude API, custom MCP servers, Python hooks, a markdown instruction tree.

BotBytes OS is my own product, not client work.