Agent Taxonomy — AI skill for Claude Code

🧬 Agent Taxonomy


**An evolutionary framework for AI agent self-improvement.**

Treat your AI agent's configuration as a living organism — with a genome that evolves through Lamarckian inheritance, horizontal gene transfer, and human selection pressure.

*"One day, frontier AI research used to be done by meat computers..."* — @karpathy, March 2026

🔬 [Take the Interactive Classifier → agent-taxonomist.dev](https://agent-taxonomist.dev/)

---

npm package scope

The zero-dependency npm package provides deterministic classification (`classify`), input validation (`validate`), taxonomy values (`VALID`), and a local interactive/demo CLI. Genome files and evolution workflows below are conventions, not functionality implemented by the package.

`validate` accepts unknown external input and reports errors without throwing for malformed JSON values. Names must be non-empty strings; optional counts must be non-negative safe integers; epithets must be strings and notable genes must be arrays of strings. Validate untrusted input before calling `classify`; classification itself is not an input-validation boundary.

The package does not implement A2A or MCP transports, discovery, registration, authentication, or payments. Its classification is descriptive metadata, not proof of agent identity, ownership, capabilities, or safety. An application can integrate these results using its own protocol adapter; taxonomy fields are not standardized A2A/MCP fields. Protocol integration requires separate implementation and t