A unified map of agent "externalization" — memory, skills, protocols, harness

A new survey ("Externalization in LLM Agents: A Unified Review of Memory, Skills, Protocols and Harness Engineering") pulls the scattered agent-engineering vocabulary into one framework: what the agent stores outside its context (memory), what reusable capabilities it loads (skills), how it talks to tools and other agents (protocols like MCP), and the surrounding execution scaffolding (harness). If you're designing an agent system and keep reinventing terms, this is a good shared reference to align a team on.

---

Sources: arXiv 2604.08224