Dissecting the Crawfish: The Anatomy of an Agent

On March 6, 2026, the opening lecture of NTU's Machine Learning course was not regression or neural networks. Hung-yi Lee chose to "dissect the crawfish" — using the open-source project OpenClaw to take apart how an AI agent actually works. When a university course with over a thousand students puts agents in week zero, the paradigm shift is complete: agents are not the frontier. They are the on-ramp.

In interviews the same season, Sega Cheng described the Silicon Valley mood: the crawfish going viral, engineers half-joking about being Claude-pilled — everyone delegating to AI. LEADAI's fourth case study, the AI Ray Lu, sits on the same technical lineage at organizational scale: a CEO's reasoning, decision criteria, and writing voice encoded into an internal agent, giving every teammate a 24-hour partner for strategic debate.

Advanced Building (Develop the Tool) is not about model size but about who gets integrated: domain experts define what is correct; end users define what is usable. Dissecting the crawfish teaches you how it works. Raising your own forces you to answer whom it lives for.

Understanding an agent is knowledge. Raising the right one is judgment.