Agent Gateway Series (Part 2 of 7) | Service Registry for the Agentic Era

In a microservices architecture, we use DNS to find services. If I need the billing service, I call billing.svc.cluster.local. It’s simple because the relationship is static: Service A calls Service B.

In an Agentic Architecture, relationships are dynamic. You don't want your "Manager Agent" to hardcode a connection to finance-agent-v1. You want it to ask: "Find me a capability that can query Q3 revenue."

Sometimes that capability is a Tool (a direct SQL query via an MCP Server).

Sometimes that capability is another Agent (a reasoning engine that analyzes the SQL output).

To solve this, TrueFoundry introduces the Unified MCP & Agent Registry. It is a single catalog that treats Tools and Agents as discoverable, interoperable assets.

1. The Unified Abstraction Layer

The registry acts as a unifying abstraction layer. It ingests disparate resources—Python scripts, LangGraph agents, Postgres Databases, SaaS APIs—and projects them all as standardized Capabilities.

To the orchestrator, a "Database" and a "Junior Analyst Agent" look identical: they are both just endpoints that accept a JSON schema and return a result.

Fig 1: Agent Registry Entry Data Model

‍

2. The "Agent-as-a-Tool" Paradigm (via MCP)

The biggest barrier to multi-agent collaboration is the interface mismatch. A LangGraph agent speaks "State Updates"; a CrewAI agent speaks "Tasks."

TrueFoundry standardizes this using the Model Context Protocol (MCP).

When you register an agent with the Gateway, the Registry automatically wraps it in an MCP Interface. This means your sophisticated "Financial Analyst Agent" looks exactly like a standard function call to the rest of the system.

• Internal Reality: A complex Python agent with memory, planning, and 3 sub-agents.
• External Interface: An MCP Tool definition.

This allows a "Manager Agent" to invoke a "Sub-Agent" just as easily as it invokes a Calculator—simplifying orchestration down to a single protocol.

Fig 2: Agent-as-a-tool Example

‍

3. Semantic Discovery: Finding Capability, Not Just Identity

How does an agent know which tool to use? It shouldn't need to know the name.

The Registry uses Vector Embeddings to map natural language intent to technical capabilities.

• Agent Intent: "I need to check if the server is healthy."
• Registry Query: Scans the descriptions of all registered MCP Servers and Agents.
• Semantic Match:
  
  • prometheus-mcp (Tool): "Query metrics from Prometheus" (Score: 0.92)
  • sre-bot-v1 (Agent): "Diagnose and fix infrastructure issues" (Score: 0.88)

The Gateway returns both options. The calling agent can then decide: Do I need a raw metric (Tool), or a diagnosis (Agent)?

4. Trust-Based Routing: "Get Me the Verified Expert"

In a decentralized enterprise, you might have five different "Coding Agents" deployed by different teams. Which one should you use?

A static registry would just list them all. The TrueFoundry Registry is Stateful. It connects to the AI Gateway's Observability Layer to track the live performance of every asset.

• coder-agent:v1 (Success Rate: 88%, Latency: 2s)
• coder-agent:v2-canary (Success Rate: 95%, Latency: 5s)

You can configure Routing Policies directly in the Registry:

"For production code changes, ONLY route to agents with >90% Faithfulness scores on the latest Eval run."

This ensures that your orchestration layer isn't just connecting to available agents, but to competent ones.

5. Topology Control: The Org Chart Firewall

As you scale to hundreds of agents and thousands of MCP tools, you risk a "Spaghetti Mesh" where every agent can call every tool. This is a security nightmare.

The Registry enforces Graph Topology. It acts as the "Org Chart" for your digital workforce.

• Rule: The Public-Support-Bot is allowed to call the Documentation-MCP.
• Rule: The Public-Support-Bot is DENIED access to the Payroll-Agent.

This check happens at the Discovery Layer. When the Support Bot asks "Who can help me with payroll?", the Registry simply returns "No results found." You cannot attack what you cannot see.

Conclusion

The Unified MCP & Agent Registry is the backbone of the Cognitive Enterprise. By treating Agents and Tools as equal, discoverable citizens—and wrapping them in a unified security and observability layer—TrueFoundry allows you to compose complex, reliable workflows from disparate building blocks.