SDK code mode shows SotA accuracy and performance for agents using APIs

Model Context Protocol (MCP) has gained popularity for standardizing tool calling, but challenges such as accuracy and token efficiency have limited its widespread adoption in AI agents composing complex API calls. To address these issues, a technique called "SDK code mode" has shown promising results by using API-specific SDKs to generate integration code, thus enhancing the accuracy and reducing the number of interactions needed to achieve desired outcomes. This method enables agents to write and execute code for complex API interactions more efficiently, leveraging idiomatic SDKs, type checking, and error messages to improve performance. Evaluations with the Increase Banking API demonstrated that SDK code mode outperformed other configurations in completeness and efficiency, though factuality remains an area for improvement. The findings suggest that with the right tools, documentation, and SDKs, MCP can be an effective means of achieving high accuracy and utility in API integrations.