Topic

From “Using Agents” to “Organization-Level Productivity”: NetEase King Crab’s Multi-Agent Design and Practice

As Agents evolve from standalone assistants to real-world business collaboration, the challenge is no longer simply whether the model is intelligent enough. The key question is how to enable multiple Agents to operate effectively in an enterprise environment—ensuring clear division of responsibilities without losing control, traceable collaboration, verifiable results, and the continuous accumulation of experience and knowledge. Using NetEase King Crab as a case study, this talk introduces a multi-Agent architecture designed for enterprise scenarios. A central orchestrator Agent is responsible for understanding objectives, generating task graphs, coordinating dependencies, and consolidating results. Domain-specific sub-Agents then execute specific tasks as their sole owners. Combined with Skills, contextual memory, permission governance, observability, and an evaluation feedback loop, these components form an Agent collaboration platform that is operable and scalable. The core idea is simple: Agents provide local intelligence, while the platform maintains global order. The value of NetEase King Crab lies not simply in adding more Agents, but in transforming AI capabilities into organizational assets that are reusable, governable, and capable of continuous evolution.