W42 - Cognitive Iteration on Agents and LLMs

I've recently been investing time in learning Agentic technologies and have gained some new insights into large models.

First, I recommend this pieceGetting AI to Work in Complex Codebases, after exploring AI coding for a long time, I view it as a programmatic overview that can guide a team to gradually form a shared understanding.

The importance of Agentic technology is becoming clear. Its value lies in enabling us to truly harness a large model's “intelligence,” transforming it from being “omniscient” to being “omnipotent.” We should distinguish two directions: one is “building Agents,” the other is “using Agents.” The boundary between them may blur over time, and to do something well you need both building and using together, but it's still important to tell whether the discussion is about construction or usage. In AI Coding, for example, I mostly see “using Agents” rather than building new intelligences.

My understanding of the MCP protocol has evolved. I originally thought of it like RPC, an application-layer protocol for service calls. Now I lean toward seeing it as more like TCP — a transport-layer protocol. MCP largely mimics traditional service-call paradigms rather than being born for multi-Agent collaboration. Real applications should sit on top of MCP and, through decomposition, scheduling, and orchestration among Agents, construct complex strategy relationships. That's how you maximize model capabilities. Agent-to-Agent interaction shouldn't be just simple calls; it needs richer protocol semantics, like HTTP layered on TCP, supporting things such as negotiated caching strategies. Therefore, what deserves attention is a new layer of protocols above MCP — for example, the emerging Agent2Agent work.

Don't treat large models as mere tools. The brain is not a hard drive but a mesh-like neural network where outputs are generated in real time. Large models are similar: not databases but parameter weights learned through training. In crude terms, larger models tend to exhibit higher intelligence. Our past R&D and delivery processes were designed around humans as the center; now we need to rethink development structure and division of labor with large models at the core. The key change is that human leverage has shifted. We should concentrate our attention on research and planning — the two highest-leverage stages — and move the standard from “human maintainability” to “Agent cooperability.”

At the same time, acknowledge the limitations of large models. Sutton still warns us that we haven't escaped The Bitter Lesson: large models do not possess intelligence in the true sense. They cannot autonomously explore unknown environments like a squirrel, nor can they generate knowledge beyond their input domain. What they can do is demonstrate planning, reasoning, and decision-making within an existing knowledge domain.