Podcast Atlas
concept Updated 2026-08-08 Tags: Ai, Agents, Workflow, Human-Agent-Collaboration

Human As Agent Tool / 人作为 AI 工具

Human as agent tool is the workflow pattern [[LiuKefan|刘可凡]] describes in EP119 对话刘可凡：用 try-catch-finally，给独立做产品的内耗写个处理流程 🐛 when he says the direction may move from “people use AI” toward “AI uses people.” In his small experiment, he built an MCP server and connected it to Claude Code so the model could call on a human to perform small real-world actions, like opening a short-video app or finding a podcast, in the same broad pattern as tool calling.

The concept is not a settled claim that AI should command people. In the source, it is a low-cost experiment inside a concrete business workflow: can an agent coordinate enough small human perception or action to improve efficiency? That keeps the idea connected to Agentic Workflow, Model Context Protocol, and Human-Agent Collaboration, while preserving Human Judgment Under AI around permission, interest alignment, and task boundaries.

Key Claims

  • Some useful agent workflows may need human hands, eyes, accounts, or local judgment when no clean API is available.
  • Treating the human as callable capability can increase efficiency, but it also raises permission, incentive, dignity, and responsibility questions.
  • The pattern is strongest when the human action is narrow, reversible, and reviewed rather than open-ended obedience.
  • A human-callable workflow should expose boundaries, expected output, and escalation rules just as ordinary tools should expose permissions and errors.
  • The source frames the idea as an experiment to validate workflow efficiency, not as proof that a broad AI-led organization model has already arrived.

Connections