Tool calling + MCP is the next moat

LLM agents that have access to MCPs & are able to tool call proficiently appear to be significantly more accurate than just using RAG to scaffold them.

After having built my own MCP servers, the output has been far better than most.

The only issue is that most models aren't great at multi-chain tool-calling, save for Anthropic's Opus. Perhaps the next step is to post-train smaller LLM agents to be specifically proficient with specific MCP servers.