Anthropic is really a chain reaction, with both theory and practice. When issuing skills, it proposed two methods to address tool bloat and token waste: layered loading/reuse of prompts, and executing code & chaining APIs/MCP (Manus calls this context offloading). The day before yesterday, while releasing Opus, these two methods were fixed at the inference API level, Tool Search Tool, solving tool discovery & lazy loading, and Programmatic Tool Calling to implement code execution tools. It feels like in the future, the Anthropic API protocol 😂 has great potential to replace OpenAI.