Uncharted: MCP and AI Agents

Hidden Shoals

We’ve never been here before. Think about it—language models are an astonishing technology adopted at an unprecedented rate by companies seeking a competitive edge. Yet, here we are, with no 'time tested' standards for production operations.

These waters are uncharted. Navigating AI Agent deployment in core production systems is challenging. Companies have deployed ‘spot solutions’—narrow, targeted applications adding specific value—but deploying transformative AI applications that are maintainable and extensible requires established standards, frameworks, and architectures. Without these, businesses risk false starts, high costs, and stranded assets hidden beneath the surface.

Anthropic introduced the Model Context Protocol (MCP), a promising approach for managing static elements of AI Agency in production. We recently shared our positive experiences working with MCP servers. Though MCP is a strong foundation, more refinement is needed.

Google now enters the fray with its Agent-to-Agent (A2A) interoperability protocol. Despite backing from major enterprises like Salesforce, Atlassian, and SAP, A2A, while a helpful contribution, does not significantly move the needle.

Google positions A2A as complementary to MCP—handling agent coordination while MCP manages context. However, our evaluation at Strategic Machines indicates A2A offers minimal innovation. It primarily repackages familiar tech stacks—HTTP, JSON-RPC, and SSE—without addressing deeper issues inherent in large-scale agent ecosystems.

Deploying agents at production scale remains genuinely unexplored territory. Security, maintainability, real-time state management, and robust performance metrics are critical yet largely unresolved challenges. Agent systems are more than typical software applications—they are dynamic entities managing sophisticated interactions and continuous updates across diverse environments. Current protocols, Google's included, barely touch upon these complexities.

We believe the future demands deeper architectural thinking. Effective state management, essential for complex interactions, should reside within the applications, providing dynamic resource discovery and session-driven actions. MCPs should serve as dedicated servers, efficiently and securely managing resources and tools for agents.

To truly advance, protocols must evolve beyond basic interoperability. Real-world agent deployments demand granular security, adaptive maintainability, precise state control, and uncompromised performance.

Navigating these seas is exhilarating yet daunting. We remain committed to charting these uncharted waters, guiding enterprises toward robust, scalable AI agent ecosystems. The seas of innovation favor those who sail boldly and prepare meticulously. The voyage into large-scale AI Agency for enterprise applications is just beginning. Reach out, and let’s explore the potential of AI Agency together.