China Proposes Thresholds for AI Agents: 20 Subtasks, 60 Agents, 200 Tools

⚡ The gist in 5 seconds - Chinese authors have proposed three numerical thresholds for AI agents: 20 subtasks, 60 agents, 200 tools. - This is about a scale metric: how many steps an agent takes, how many agents work in tandem, and how many tools are kept in the loop. - There is still no common standard for comparing agents — the breakdown appeared in an article on Habr. ### 🔍 What was found The article on Habr examines the Chinese proposal with specific numbers: 20 subtasks, 60 agents, 200 tools. It is an attempt to set measurable boundaries of "agent capability" — that is, to describe the complexity of a task an agent can handle through three countable parameters rather than vague talk about "smartness." The author questions whether such a scale works in practice: is it valid to compare two agents with a single number. The mechanics of the debate are simple. A chain of 20 subtasks tests whether the model can hold a long plan. 60 agents is about multi-agent architecture, where several executors split the work. 200 tools is about the breadth of the API park an agent gets access to. Each parameter measures its own thing, and collapsing them into a single rating without losses is the main problem. ### 💡 Why it matters Agents are moving into production tasks, yet there is still no way to compare them: public benchmarks answer narrow questions and transfer poorly to real-world scenarios with long chains of calls. Numerical thresholds are an attempt to provide at least a rough scale for choosing a model for a specific workload. For now, this is a proposal, not a standard: how — and whether — it will be adopted does not follow from the material.