It's based on being able to rebuild the system prompt on every turn with the skills the agent needs. So an agent can have lots of skills without overwhelming the context window. It keeps the input tokens down, and with smaller system prompts could help reduce latency. https://claude.ai/share/585c872c-2584-49a3-8a9c-509d6cf16f94 https://github.com/vercel-labs/agent-skills