Microsoft Pushes for Distributed Skills Over MCP to Boost Agent Architecture

Microsoft's Agent Framework team has proposed a new pattern for agent architecture that replaces specialist agents with skills served directly over the Model Context Protocol (MCP). The proposal, outlined in a blog post by Tommaso Stocchi on September 16, aims to reduce latency and improve scalability. In the traditional Agent-to-Agent (A2A) architecture, each specialist agent runs its own model to interpret requests, select tools, and write responses.

Under the new distributed skills over MCP pattern, specialist agents publish a description of their capabilities, along with typed MCP tools, which are then loaded by the parent agent on demand. This approach reduces the number of model calls required for each request, resulting in a 60% decrease in latency. However, it also increases the token count, with the skills path consuming around 22% more tokens than the A2A path.

The team has built a demo ski resort advisor application that uses both architectures side by side, demonstrating the performance benefits of distributed skills over MCP. While the proposal is experimental and subject to change, Microsoft plans to standardize this approach through the SEP-2640 extension to the MCP protocol.