This position paper argues that static tool-registry protocols, most prominently the Model Context Protocol (MCP), are unsuitable as foundational integration architectures for production-grade agentic AI in heterogeneous enterprise environments, and that the field should converge on “Real-Time Discovery and (Self) Coding” (RTDC) Integration as the necessary architectural alternative. MCP was introduced to standardize connections between language models and external systems through a uniform JSON-RPC 2.0 client-server interface. Despite rapid adoption in developer tooling, we demonstrate that MCP doubles the enterprise integration maintenance and vulnerability surfaces, fails categorically against legacy systems that harbor the most critical enterprise data, exhausts language model context windows through static tool enumeration at production scale, induces selection collapse in large tool registries, and introduces security vulnerabilities, including tool poisoning, prompt injection, and supply chain compromise. These are fundamentally incompatible with enterprise zero-trust architectures. We survey three alternative paradigms: user-interface automation, real-time context lake architectures, and terminal agents. We show that each resolves only a subset of the enterprise integration challenge. We then argue for RTDC Integration: a paradigm in which autonomous meta-agents dynamically discover system interfaces, and schemas, at runtime, synthesize and validate integration code in sandboxed execution environments, and accumulate versioned, reusable capability artifacts through a continuous discovery-synthesize-verify-promote loop. Unlike static protocol registries, RTDC integrates with legacy mainframes, undocumented APIs, and proprietary systems without prior engineering effort, and enables the eventual decommissioning of those systems through autonomous logic internalization. This is a capability no protocol-based approach can provide.
Stephane Maes (Sun,) studied this question.