Willie Lin | Smart Manufacturing, Agentic AI & Decision Intelligence

Pip: Willie Lin writes about smart manufacturing and agentic AI — territory where the vocabulary moves faster than most people’s ability to keep up, which is either exciting or exhausting depending on your Friday afternoon.

Mara: Today we’re covering the architectural building blocks of AI agents: what they can do, how they connect to the systems around them, and why both questions matter for industrial applications.

Pip: Let’s start with the distinction between Agent Skill and Model Context Protocol — and why conflating them is a real design problem.

Agent Skill vs. Model Context Protocol

Mara: The core question here is how an AI agent actually gets useful work done — and it turns out the answer splits into two separate concerns that are easy to confuse.

Pip: The post draws the line cleanly: “Agent Skill defines what an AI agent is capable of doing, while Model Context Protocol defines how the agent connects to external tools, data, and systems.”

Mara: So the upshot is: capability and connectivity are distinct layers. You can have a well-designed skill and still be unable to act on it without the connection infrastructure underneath.

Pip: The manufacturing example makes this concrete. Someone asks why production yield dropped today. The skill tells the agent how to reason about yield loss — check defect categories, equipment status, process parameters. That’s the analytical logic.

Mara: But the agent still needs to actually reach the MES production records, IoT sensor data, ERP order information, and quality inspection reports. That retrieval is what MCP handles — it’s the bridge to the real operational environment.

Pip: Which means building a reliable agent isn’t just prompt engineering. The post is explicit that you need clear task capabilities, trusted data access, and traceable recommendations — not one of those alone.

Mara: Right, and the framing the post uses is useful here: Agent Skill makes the agent useful for a specific task, while MCP makes the agent practical in a real system environment. Those are two different problems requiring two different design decisions.

Pip: The analogy in the post is tidy — skill is the agent’s professional expertise, MCP is the bridge to the world. Separate things, both required.

Mara: And that distinction becomes especially sharp in industrial settings, where the systems an agent needs to reach — MES, ERP, IoT platforms, SOPs — are numerous, heterogeneous, and already exist before any AI layer arrives.

---

Pip: So the architecture question isn’t just “what can this agent do” — it’s whether the connection layer is actually there to back the capability up.

Mara: Both halves have to be designed deliberately. That’s the thread worth carrying into whatever comes next.