@Yuancheng: ➤ New ideas and practices for Agent Harness are still emerging. Lately I came across **OpenSquilla**, an open-source, locally-hosted AI Agent. ① It features intelligent model routing—for the same task, token cost is 60-80% less than OpenClaw …

➤ New Agent Harness ideas and practices continue to emerge. Recently I saw **OpenSquilla**, an open-source, locally-hosted AI Agent. ① It has intelligent model routing—for the same task, token cost is 60-80% less than OpenClaw, and when running tasks there's a "slot machine" animation telling you how much money you saved on this task. Routing saves more than just money: each model has its strengths and different prices, and benchmark performance often doesn't match performance on specific tasks. Rather than obsessing over a single best model, it's better to assign tasks to the most suitable one based on the task—what it truly optimizes is "how much intelligence you get per dollar of tokens." ② They just released MetaSkill. Simply put, it's the "skill of skills"—a meta-markdown that tells the model how to retrieve, filter, and combine atomic Skills. Because a single Skill only does one thing: search is search, document is document. If the user acts as the dispatcher, always thinking about which Skill to use next, it becomes chaotic when there are many Skills. MetaSkill takes over the "project manager" role—it decides which steps can be parallel and which must be sequential, allowing the Agent to learn to organize skills on its own. ③ There's also meta-skill-creator: a one-sentence requirement automatically synthesizes a new meta-skill, compressing a 30-minute manual task into 3 minutes. It's like the Agent creating skills by itself. ➤ Why now? Three things have converged: models can now understand complex multi-step orchestration instructions; the community's Skills are exploding in number, to the point where a higher abstraction layer is needed to filter them; large model online trial-and-error is still too expensive, so optimization must be moved up to the Skill organization layer. These three threads point to the same thing—the problem that Agents need to solve is shifting from "whether they can call tools" to "whether they can organize tools." **➤ Going deeper:** Could the Harness layer be just a transitional concept, eventually eaten by stronger models? I don't think so—as long as models continue to diversify and skills and tools keep growing exponentially, "how to organize and how to schedule" will remain a problem that won't disappear. Very interesting direction. GitHub: