Morgan Stanley Rethinks API Strategy to Support Agent-First Architectures

Morgan Stanley is fundamentally reconsidering how it designs and deploys APIs in response to the rapid emergence of AI agents that need to discover and interact with business services. Rather than treating APIs as endpoints for traditional integrations, the firm is adopting a dual approach: pairing the Model Context Protocol (MCP) with CALM (Architecture as Code) to create platforms that agents can understand and orchestrate.

The shift reflects a broader industry transformation. A year ago, MCP was barely on the radar. Today, major platforms including OpenAI, GitHub, and others have embraced it as a standard for connecting language models to tools and data sources. For Morgan Stanley, this timing prompted a comprehensive rethink of its API program strategy.

The Gap Between APIs and Agent Interfaces#

Traditional API specifications like OpenAPI have long documented what services can do, yet they don't always translate well to agent-driven workflows. MCP introduces three key elements beyond standard API descriptions:

• Tools: Structured operations that agents can invoke, essentially composing multiple API calls into reusable, business-meaningful actions
• Prompts: Reusable, parameterized instructions that guide agent behavior in consistent patterns
• Resources: Documents and contextual data that agents can access to make better decisions

The business demand is clear: executives and technical teams want to ask questions in plain English—"Show me my top 10 Vodafone trades"—and have agents figure out which systems to query, how to interpret ambiguous inputs, and assemble answers without manual intervention.

Where Complexity Emerges#

While the MCP protocol itself is simple (essentially client-server communication), managing agents at scale introduces new challenges:

• Tool disambiguation: As the number of available tools grows, overlapping descriptions and similar-sounding operations force agents to ask clarifying questions, increasing token costs
• Chatty interactions: Agents may need to validate assumptions multiple times, driving up API call volumes and expenses
• Governance gaps: Traditional API gateways weren't designed to understand agent intent or enforce business rules in real-time agent orchestration

This complexity has led to the emergence of specialized control planes—intermediary layers that sit between the protocol and agents, enforcing policies without embedding business logic in the gateway itself.