A new formal analysis finds that the Model Context Protocol has structural gaps compared to an older research framework for agent-tool integration.
Researchers published the first process calculus formalization of both Schema-Guided Dialogue (SGD) and MCP, the industry standard that lets AI agents call external tools. Using a mathematical framework called bisimulation, they proved the two protocols are equivalent in one direction - SGD can map cleanly to MCP - but that the reverse mapping is partial and lossy. In plain terms: MCP cannot fully express everything SGD can. The team identified four properties it says are necessary and sufficient for full equivalence: semantic completeness, explicit action boundaries, failure mode documentation, and inter-tool relationship declarations. They call the extended spec MCP+.
This matters because MCP has become the de facto plumbing layer for agentic AI systems, and "plumbing" with expressiveness holes creates verification problems that compound as agents grow more autonomous. If tool schemas cannot capture all the behaviors an agent might invoke, formal safety guarantees become impossible to make - which is a real problem for anyone deploying agents in high-stakes settings.
MCP is an industry-backed spec that has moved fast and attracted wide adoption; SGD is a research artifact that never faced the same deployment pressure. That the research framework turns out to be the more expressive one is a familiar story in software standards - rigor often loses to momentum, and then someone has to write a paper explaining what got left out.