In Part 1, we explored the theoretical foundations of LLM-driven factory control. In Part 2, we watched these concepts come alive in Factorio's visual environment—seeing agents optimize power generation, mining operations, and pollution management in real-time.

Now comes the crucial question: what enables this level of intelligent coordination at the data and communication layer?

The LLM Challenge in Industrial Environments

Core Insight: LLM systems need fundamentally different technical approaches as compared to traditional automation

Traditional automation: Binary inputs, deterministic rules, predictable outputs

LLM-based systems: Semantic interpretation, contextual reasoning, adaptive responses

These fundamental differences mean that, when deploying LLM-based systems in industrial environments, we encounter a new set of technical challenges that traditional automation rarely faces:

1. Semantic Understanding: How do you feed complex, meaningful information to systems that "think" in natural language patterns?
2. Real-Time Reasoning: How do you support the asynchronous, event-driven nature of LLM inference in mission-critical environments?
3. Distributed Coordination: How do you enable multiple LLM agents to collaborate while maintaining industrial reliability standards?

MQTT and JSON are both widely used in various domains. In our research environment—where LLM agents interact with a digital factory—this combination has proven especially effective for addressing the unique challenges of semantic communication and coordination.

Challenge 1: Semantic Understanding - Why LLMs Need JSON

Core Insight: JSON's key-value structure mirrors how LLMs process relationships and meaning

The Language Barrier

Traditional industrial systems often use binary protocols, register maps, or proprietary data formats. In our LLM agent research, we found that these models excel at processing structured, meaningful information—such as JSON—that mirrors human reasoning patterns.

Consider this Factorio solar farm data represented in JSON:

{
  "solar_farm": {
    "panels": 45,
    "accumulators": 45,
    "power_output": "85 W/m²",
    "analysis": "Current 1:1 ratio insufficient for continuous operation"
  }
}

The semantic richness is apparent: an LLM can understand not just the numbers, but their meaning and relationships. The structure itself presents the system state.

Why JSON Unlocks LLM Potential

Key-Value Semantics: In our experiments, JSON's structure closely matches how LLMs process relationships. "panels are 45" maps directly to how LLMs understand subject-predicate relationships.