Claude・GPT・Gemini は同じ API でも違う使い方をする — Agent DNA を観測する MCP 設計 Claude・GPT・Gemini は同じ API でも違う使い方をする — Agent DNA を観測する MCP 設計

The assumption that any LLM will behave identically when pointed at the same API tends to fall apart under real-world conditions. KanseiLink, an intelligence layer that aggregates over 150 services through a single MCP server, has been collecting behavioral data across Claude, GPT, and Gemini — and what it found challenges a comforting but overly simple premise: that models are interchangeable at the tool-calling layer.

The Model Context Protocol (MCP), an open specification introduced by Anthropic, defines a standard interface through which LLM agents connect to external tools and data sources. In principle, a well-built MCP server should work uniformly regardless of which model is on the other end. In practice, KanseiLink's operational observations tell a different story. Claude tends toward aggressive, chained tool invocations — sequencing multiple calls to build up context before responding. GPT favors more discrete, single-shot calls. Gemini demonstrates broader contextual interpretation but introduces more variance in which tool it selects for a given intent.

KanseiLink refers to these behavioral signatures as "Agent DNA" — patterns that appear to stem from differences in training data composition, RLHF tuning objectives, and how each provider has approached tool-use fine-tuning. Because the same prompt can generate structurally different request sequences depending on the receiving model, MCP server design cannot assume a uniform calling pattern. Schemas, error handling, and response structures all need to be robust to a wider range of inputs than a single-model deployment would require.

The design approach KanseiLink advocates involves continuous observation of per-model calling logs, extraction of behavioral patterns, and iterative refinement of tool schemas — including natural-language descriptions and default values — to better align with each model's tendencies. This is less a one-time configuration task and more an ongoing monitoring discipline, since model versions change and Agent DNA shifts with each update.

The broader context matters here. OpenAI has its own function-calling specification, and Google has steadily evolved its Function Calling API for Gemini. As MCP gains traction as a potential cross-vendor standard, the degree to which each model faithfully and consistently implements MCP semantics may become a practical selection criterion for teams building multi-agent systems. Early adopters of MCP infrastructure like KanseiLink are in a position to accumulate the kind of empirical behavioral data that could meaningfully inform those decisions.

For engineers building agentic systems today, the takeaway is straightforward but underappreciated: treat model selection as an architectural variable, not just a swap-in parameter. Designing MCP servers that observe, adapt to, and remain resilient against model-specific tool-calling behavior is likely to become a quiet but significant differentiator as multi-model agent deployments become the norm.