Anthropic just launched a "sub-agents" feature for Claude Code, which lets you define specialized agents in Markdown files that the main agent can delegate tasks to. This is interesting to me because a month ago, I launched an experimental project called LLMunix. The core idea was to build an OS-like framework for an LLM where agents and tools are defined as simple, version-controllable Markdown files. The system would achieve a user's goal by creating a plan and then orchestrating these components. A key part of the architecture was delegation. Since Claude didn't have this natively, I built a run_agent tool. It's essentially a shell script that takes the path to an agent's Markdown file and some JSON arguments, then spawns a separate, non-interactive Gemini CLI or Claude Code sub-process to execute it. It was a hack, but it worked to prove the concept of a higher-level orchestrator using specialized, file-based agents. The new native implementation is strikingly similar: Agents are defined in Markdown files with YAML frontmatter. They live in a specific directory (.claude/agents/). The orchestrator (Claude Code itself) delegates to them based on their description. The native version is obviously superior – it offers true context isolation, proper tool permissioning, and is much cleaner than my sub-process hack. I'm now planning to refactor LLMunix to use these native sub-agents directly. It will simplify the codebase significantly and make the project a much better example of this emerging architectural pattern. To me, this confirms that the future of complex AI work isn't a single monolithic model, but an orchestrator delegating tasks to a collection of specialized, human-readable components. It's a fun bit of parallel evolution, and I'm excited to build on top of the real thing now.

Curious to hear what others think of this multi-agent, file-based approach.