Rethinking API Programs for AI Agents with MCP and Architecture as Code
This presentation from Morgan Stanley explores how traditional API programs are evolving to support AI agents, introducing the Model Context Protocol (MCP) as a key enabler. It details the use of Architecture as Code via the CALM framework for automating API deployments and enforcing governance, alongside strategies for scaling enterprise AI workflows with zero-downtime platform upgrades. The discussion highlights the architectural shift required to manage agent-to-agent communication and the complexities introduced by increased tool orchestration.
Read original on InfoQ CloudThe Shift to APIs for AI Agents (MCP Era)
The article discusses a significant industry shift towards APIs designed for AI agents, driven by the Model Context Protocol (MCP). MCP is an open protocol connecting LLM-based applications to tools and data, facilitating discovery, invocation, and validation. Unlike traditional OpenAPI specifications, MCP aims to provide a more business-relatable and developer-friendly way to expose capabilities, often leading to rapid adoption due to its ease of integration with LLMs.
Key Components of Model Context Protocol (MCP)
MCP differs from standard API specifications by introducing concepts like 'tools' (structured operations combining API endpoints for specific tasks), 'prompts' (reusable, parameterized instructions for guiding agent behavior), and 'resources' (documents and data providing context to agents). This richer context helps agents orchestrate complex workflows.
Complexity in Agent Orchestration and Tool Selection
While exposing individual MCP tools is simple, the complexity rapidly increases with the number of tools available. Overlapping definitions, ambiguity in natural language descriptions, and the 'chatty' nature of agent interactions drive up token costs and necessitate specialized gateways and control planes. This means API gateways, which traditionally avoid business logic, now need to be more aware of interaction design to effectively manage MCP-based agent communications and disambiguate tool usage.
Architecture as Code with CALM for API Programs
Morgan Stanley leverages CALM (Architecture as Code) to modernize its API program, enabling rapid deployment and automated governance. CALM provides a core modeling capability (JSON schema for boxes and arrows), CLI tooling, and powerful patterns that enable reuse and enforce guardrails. This approach allows deploying numerous APIs based on a few standardized patterns, integrating security and continuous deployments effectively. CALM Hub serves as an Artifactory for architectures, patterns, and controls.
- Core Model: JSON schema for defining architectural components and their relationships, with typed information for enhanced specificity.
- CLI Tooling: Command-line interface for interacting with and managing CALM definitions.
- Patterns: Reusable architectural blueprints that standardize deployments and accelerate development, enforcing best practices.
- CALM Hub: A central repository for architectural models, patterns, and controls, facilitating enterprise-grade management.
The integration of Architecture as Code with deployment gates and zero-downtime platform upgrades is critical for safely scaling enterprise AI and agentic workflows, ensuring automated governance and security are embedded throughout the lifecycle.