Manus Acquisition Shows: Agent Engineering Outvalues Model Training

Manus Acquisition Sparks Technical Reflections

The recent news that Manus was acquired by Meta for billions of dollars has triggered widespread discussion: how can a company that does not train foundation models, but only engineers the periphery, command such a valuation? This gives confidence to builders of LLM applications, highlighting the gap between base model capabilities and final user needs. Someone must package the engine into a luxury car that users can actually drive. The author references a Zhihu discussion (https://www.zhihu.com/question/1989243102427378443) where both sides present arguments.

Core Engineering Challenges of the Last Mile

The author identifies three fundamental problems that agent engineering must solve:

Memory Engineering : LLMs have no native memory. They depend on external agent-framework memory modules to store, process, and retrieve past information, thereby helping users organize context more effectively.

Tool Safety Sandbox : Agents need a secure, user-familiar operating environment. Examples include Cursor as a developer workbench and the browser as a general-purpose operating platform.

Agent Framework as the Vehicle : The LLM is a non-deterministic engine; the agent framework provides the car. Memory engineering, context engineering, and tool collections (tools, MCP, skills, etc.) act as the wheels and steering wheel for that engine.

The Nature of Agent Development: Alchemy on Uncertainty

Despite the scaffolding, the work remains essentially alchemy on uncertainty. Developers try various methods to make inputs more effective, to trigger model parameters more accurately, and to attach hook functions to outputs so that matched tools are called and results aggregated. The agent framework is merely the scaffold; the author's inspection of CrewAI and OpenAI's framework reveals a shallow technology stack with no deep secrets. Many fancy concepts are proposed, but the essence is experimental tinkering on input, orchestration, and output. It is precisely this uncertainty that fuels infinite imagination.

Five Business Models for AI Companies

The author categorizes successful company models into five types:

Monopoly Resource Enterprises : Minerals, oil, banking, utilities — typically state-owned.

Internet Platform Enterprises : Tencent, Douyin, Alibaba — first-mover advantage and user-habit lock-in; Pinduoduo succeeded by integrating low-end supply chains and tapping the desire for cheap goods.

Manufacturing Companies : Phone and telecom makers — require not only technology but also supply-chain management and after-sales service; technology has plateaued, future competition is cost management.

AI Foundation Model Service Companies : The future "water companies"; tokens are water, a basic production factor for every enterprise.

AI Application Companies : Currently the tech stack is shallow, making moats hard to build; most will likely merge into existing services. A few may succeed by accumulating deep engineering experience in context management, memory management, and toolboxes.

Conclusion: Back to the Grind

Ultimately, if agents truly succeed, the work will return to fundamentals: continuous alchemy to accumulate experience, gathering valuable data, building toolboxes, and understanding user needs and habits to deliver great UX — using the agent framework to unlock the LLM's potential.