#22: Why We Built Our Own MCP Server Infrastructure (And What We Learned)

The Singapore Supply Chain Reality

When I first arrived in Singapore, you simply couldn't miss the endless stacks of containers lining the coast. It’s orchestrated chaos, the heartbeat of the city. But working in this industry, I've learned the hard way that behind every smooth shipment lies a mountain of messy data.

Six months ago, I joined Tetrixx with a mandate that sounded simple on paper but was terrifying in practice: build AI agents that could actually do useful work in this chaotic environment. We weren't interested in chatbots that write poetry. We needed agents that could navigate the labyrinth of the supply chain world.

I thought I was building an agent. I was wrong. What I didn't expect was that the most critical thing we would build wasn’t the AI itself, it was the digital plumbing that allowed the AI to survive. This is the story of Agent Forge, our multi-agent platform, and how the Model Context Protocol (MCP) completely changed how we approach AI architecture.

The "One Smart Agent" Trap

In the early days, we fell into a trap that I think 90% of AI engineers fall into: The Monolith.

When people imagine AI, they picture a single, super-intelligent "God Mode" assistant. However, in complex domains like logistics, the monolithic approach hits a ceiling quickly.

The Context Window Wall: As input files grow larger, a single agent begins "forgetting" instructions from the start of the prompt.

The Logic Tangle: Asking one model to switch between being a data parser, a logic engine, and a pricing analyst results in hallucinations.

We realized we needed a way to build specialized teams, not just single bots. This led us to build Agent Forge.

Agent Forge isn't a single tool; it is a flexible platform that allows us to spin up any number of specialized agents and wire them together. It’s not tied to one project. Whether we are parsing invoices, tracking charges, or auditing finance log, we can simply include any agent we desire on the roster.

For example, when dealing with complex document comprehension, we don't ask one bot to do it all. We use the platform to deploy a specific squad where each agent has a distinct, narrow responsibility, separating structural analysis from data extraction.

The Breakthrough - MCP and Agent Forge

This is where the Model Context Protocol (MCP) acts as the foundational layer.

Think of MCP as a standard interface between AI models and their tools. We designed Agent Forge to sit on top of this protocol, making connectivity effortless. In our platform, creating a new agent is as simple as defining its role and plugging it into an MCP server.

The true power of the architecture lies in its modularity and orchestration:

Plug-and-Play Connectivity: We can spin up multiple agents and connect them easily to any MCP server we have running, whether it's for file processing, database lookups, or API calls. The agents simply appear as capable entities within the system

Seamless Collaboration: Because the platform handles the wiring, agents can collaborate easily. They aren't isolated scripts; they are nodes in a network that can pass tasks back and forth.

The Multiplier Effect

The combination of Agent Forge and MCP turns AI development from a custom coding challenge into a modular assembly line.

Write Once, Use Everywhere: Once a capability is standardized as an MCP server, it becomes universally reusable. A server built for data extraction in one project can be instantly leveraged by a completely different team for a different purpose. The server doesn’t care who asks; it just provides the standard interface.

Model Agnosticism: The platform allows for swapping the "brain" of any agent instantly. High-volume structural agents can run on faster, lightweight models, while complex reasoning agents utilize more powerful architectures. The infrastructure handles the handoff seamlessly, optimizing both cost and performance.

Context Importing: If an agent encounters a limit or needs to transfer work, the current state is preserved. A different agent can then explicitly import the context from the previous one to continue exactly where it stopped, or simply use its output as a starting point.

What We Learned

Looking back at the last six months at Tetrixx, three lessons stand out clearly for anyone trying to build efficient ai agents:

Specialization beats Generalization: A team of small, focused agents will always outperform one massive model trying to do it all. Specialized agents are easier to debug, cheaper to run, and significantly more reliable.

Collaboration is the Force Multiplier: Collaborative frameworks are the true secret sauce, enabling agents to build on each other's knowledge and collective progress.

Infrastructure is an Investment, Not Overhead: We spent months building the Agent Forge platform. That felt risky at the time. Today, that investment means we gain hours of efficiency with every new addition.

The future of AI agents in the supply chain isn't a single super-computer running the port. It’s teams of specialized digital agents, communicating via standards like MCP, working alongside humans to tame the chaos.