MCP Won. Now What?

Inside the standardization gap holding back agentic AI

Two and a half million years ago, our ancestors began chipping stones into cutting tools. This wasn't just a technological achievement—it was a cognitive one. Research in evolutionary neuroscience has demonstrated that the social transmission of Oldowan tool-making technology enhanced the development of teaching and language, creating a co-evolutionary dynamic that lasted millions of years. Tools didn't just extend human capabilities; they shaped human cognition itself.

This insight carries profound implications for artificial intelligence. We've spent the last decade scaling models—more parameters, more compute, more data. But if the human precedent holds, the path to more capable AI doesn't run solely through bigger models. It runs through better tools.

The Tool-Intelligence Feedback Loop

The archaeological record shows a clear pattern: as tool complexity increased, so did the cognitive demands of making and using them. Early Oldowan tools required basic motor coordination. Later Acheulean tools demanded hierarchical action organization, working memory, and the ability to plan multiple steps ahead. The brain regions activated during complex tool-making overlap significantly with those used for language—suggesting that tool use and higher cognition developed together.

AI agents today face an analogous situation. Large language models possess remarkable reasoning capabilities, but they remain constrained by their isolation from the world. An LLM can discuss database queries eloquently but cannot execute one. It can plan a workflow but cannot act on it. Tools bridge this gap—and like our ancestors, AI agents will become more capable as their tools become more sophisticated.

This is why the emergence of the Model Context Protocol matters. MCP, introduced by Anthropic in late 2024, provides a standardized way for AI systems to connect with external tools and data sources. It has achieved remarkable adoption: OpenAI integrated MCP across its products in March 2025, Google DeepMind confirmed support for Gemini, and in December 2025, MCP was donated to the Agentic AI Foundation under the Linux Foundation. The protocol problem has been solved.

But a new problem has emerged.

The Schema Gap

MCP answers how agents talk to tools. It doesn't answer how tools should describe themselves.

Consider what happens when an AI agent encounters a new tool. It reads a description, examines the input schema, and decides whether and how to invoke the function. But MCP doesn't mandate how that description should be written, what metadata is required, how tools should be categorized, or how pricing information should be expressed. The result is inconsistency.

This isn't theoretical. An active proposal in the MCP repository (SEP-1382) explicitly states the problem: without standardized documentation practices, the MCP ecosystem will continue to suffer from inconsistent tool interfaces that confuse both implementers and consumers. Testing by the Mastra team found that different models handle schemas differently—OpenAI models throw errors on unsupported properties while Google Gemini models silently ignore constraints. They had to build a compatibility layer to reduce tool calling error rates from 15% to 3%.

The academic literature has formalized this as "tool hallucination"—a phenomenon where models improperly select or misuse tools, leading to erroneous task execution and increased operational costs. Researchers categorize these into tool selection hallucination (picking the wrong tool) and tool usage hallucination (providing incorrect parameters). The root cause often traces back to documentation issues: redundant information, incomplete descriptions, or lack of standardization that impairs the agent's ability to properly use tools.

When tool interfaces are inconsistent, agents struggle. They fall back to familiar patterns. They hallucinate function calls. They fail silently. The protocol works perfectly while the ecosystem remains fragmented.

Why Independent Standardization Won't Work

One might argue that the community will naturally converge on best practices. History suggests otherwise.

OpenAPI has existed since 2010—a well-structured, stable specification for describing APIs. Yet it hasn't achieved widespread adoption in the AI function-calling ecosystem. Why? Because JSON schema specifications differ between AI vendors, and each OpenAPI version has its own distinct schema specification. Without enforcement, standards become suggestions.

MCP faces the same risk. If every vendor implements schemas differently, the fragmentation just moves up a layer. We'll have a universal protocol with a thousand incompatible dialects.

This is where the human evolutionary parallel becomes instructive again. Tool-making didn't advance through individual innovation alone. It required social transmission—teaching, demonstration, shared standards passed between generations. The tools themselves became a form of standardized knowledge that accumulated across communities.

AI tooling needs the same: not just a protocol, but an ecosystem that enforces consistency while remaining open to evolution.

The Hub-and-Spoke Model

AgentPMT has built precisely this infrastructure. At its core is a simple proposition: vendors comply with a standardized schema format once, and their tools become instantly accessible to every agent in the network.

This isn't about creating a proprietary layer on top of an open protocol. The schemas AgentPMT enforces are available for anyone to examine and adopt. The value lies in enforcement and network effects. When a vendor lists a tool on the AgentPMT Marketplace, request formats are validated against vendor-provided schemas before they ever reach the agent. Non-compliant requests are rejected. Failed purchases from malformed data become rare.

The X402 Direct payment protocol adds another dimension. User-defined guardrails are stored on-chain as payment filters in a smart contract. Only the user can modify these rules. This means agents can be given economic agency without unrestricted access to funds—solving one of the thorniest problems in agentic commerce.

But the architectural insight that matters most is this: when standards evolve, all vendors automatically benefit without changing anything.

Imagine SEP-1382's documentation best practices become official MCP specification. AgentPMT can adopt them into its validation layer. Every tool already listed in the marketplace—having already been standardized—becomes compliant with the new spec. No vendor needs to update their integration. No agent needs to be reprogrammed. The hub absorbs the change and propagates it outward.

This is how rapid, far-reaching progress happens. Not through asking thousands of independent vendors to independently implement evolving standards, but through infrastructure that makes compliance automatic.

The Call to Action

The companies building AI's future—Anthropic, OpenAI, Google, Microsoft—have recognized that interoperability requires open protocols. MCP's donation to the Linux Foundation signals a commitment to shared infrastructure over proprietary lock-in.

But protocols are necessary, not sufficient. They need connective tissue: validation layers, schema standards, enforcement mechanisms, and economic infrastructure that makes compliance worthwhile.

Supporting open protocols means supporting the companies building this connective tissue. It means recognizing that the next layer of the AI stack—the layer between protocols and practical deployment—requires the same attention and investment we've given to model development.

Humans dominated not through superior physical attributes but through our ability to create, share, and accumulate tool-based knowledge across generations. AI will follow a similar path. The agents that become most capable won't just be the ones with the largest models—they'll be the ones with access to the richest, most consistent, most discoverable tool ecosystems.

We're building that ecosystem now. The decisions we make about standardization, enforcement, and openness will shape AI capability for years to come.

The tools we give AI will determine what AI becomes. Let's build them well.

AgentPMT is building infrastructure for agentic commerce, including the X402 Direct payment protocol and a marketplace for AI agent tools. Learn more at agentpmt.com.