139. 【Agent的综述】和苏煜聊Agent技术史、OpenClaw Moment、边界的消弭和社会的辐射

Summary

This 张小珺Jùn｜商业访谈录 episode interviews Su Yu / 苏煜 about the longer technical history of AI agents, from logical agents and expert systems through neural RL agents, Semantic Parsing, and modern Language Agent systems. Su uses Memory-Autonomy Framework to argue that agents have always been about adapting to complex environments, then treats OpenClaw Moment as an interaction-form shock comparable to the earlier ChatGPT shift. The episode’s forward-looking thesis is that Web, desktop, mobile, coding, and Computer Use Agent boundaries will converge into Universal Digital Agent systems whose bottlenecks are Continual Learning, World Models, reliability, speed, cost, and Specialized Intelligence.

Key Claims

• Su Yu / 苏煜 defines an agent as a bounded entity in an external environment that acts toward goals, making agency a core AI problem rather than a recent LLM label.
• Memory-Autonomy Framework separates the agent problem into memory, including semantic, episodic, and procedural knowledge, and autonomy, including perception, reasoning, decision, and action.
• Logical agents and expert systems encoded expert knowledge in formal rules and inference engines, but were limited by expression and knowledge-acquisition bottlenecks.
• Neural agents improved autonomy in repeatable environments such as games, but often compressed reasoning into a single forward pass and remained weak on explicit memory.
• Semantic Parsing expanded agents’ action space by translating natural language into machine-readable forms that could call databases, knowledge graphs, websites, or tools.
• Language Agent marks the ChatGPT-era shift where language becomes the scaffold for perception, reasoning, planning, tool use, and action.
• Chain-of-thought-style generation gives models adaptive computation by letting them spend more tokens on harder reasoning.
• Recent agent history runs through Chain of Thought, ReAct, LLM Planning, Mind2Web, Toolformer, AutoGPT, GPT-4V, MMMU, CACT, WebArena, OSWorld, and Uground.
• OpenClaw Moment is framed less as a new algorithm than as a shift in interaction form: permissions, always-on operation, local or personal context, and IM-style reachability made agent capability feel newly real.
• Universal Digital Agent is the expected convergence point: temporary labels such as Web Agent, desktop agent, mobile agent, coding agent, and computer-use agent should dissolve as agents operate across the digital world.
• Coding is described as the digital world’s underlying fabric, so coding agents can break boundaries among GUI, CLI, API, and other software surfaces without ceasing to be language agents.
• The episode qualifies pure CLI/headless enthusiasm: GUI remains important for human understanding, verification, trust, and audit, and much operational knowledge is already encoded in graphical systems.
• Specialized Intelligence is the near-term opportunity for Neo Cognition: once general intelligence gets cheaper, differentiation moves toward fast adaptation inside professions, domains, organizations, and environments.
• Continual Learning is treated as the shared bottleneck behind memory, self-learning, agent post-training, personalization, expert agents, and broader World Models.
• The major social risk is not near-term AI extinction in this account, but job displacement, distribution of gains, and whether frontier agent capability can be made broadly usable.

Key Quotes

“Memory + Autonomy” — Su’s compressed framework for comparing agent paradigms.

“OpenClaw Moment” — the episode’s name for the personal-agent interaction shock.

“Universal Digital Agent” — the future convergence target beyond separate web, coding, and computer-use agents.

Connections

• Su Yu / 苏煜 and Neo Cognition — guest and research lab behind the episode’s specialized-agent thesis.
• 张小珺Jùn｜商业访谈录 — show context for the interview.
• Language Agent, Memory-Autonomy Framework, Semantic Parsing, and Computer Use Agent — technical-history concepts added by the source.
• Open Claw, OpenClaw Moment, Open Cloud, Agent Harness, and Agent-Facing Interfaces — agent-product and interface branch extended by the source.
• Universal Digital Agent, Specialized Intelligence, Continual Learning, and World Models — future capability and bottleneck concepts emphasized by the episode.
• Persistent Agent Memory, AI Skills, Agent Post-Training, Agent RL, Agent Self-Evolution, and Model Harness Co-Evolution — existing wiki concepts the source connects to its learning-and-agent-history frame.
• Anthropic, OpenAI, Google, Nvidia, xAI, ByteDance, and Zhipu AI — major labs or companies named as moving around agent and computer-use directions.

Contradictions

• No direct contradiction with prior wiki content.
• The source qualifies the wiki’s Headless Software and Agent-Facing Interfaces thread by arguing that GUI will remain important because humans need visual verification and because existing software already stores business logic in graphical workflows.
• The source also broadens World Models beyond video, robotics, and physical simulation into professional and organizational “small worlds” that include workflow, software operation, interpersonal context, and theory of mind.