The 5 tools at 95% versus 30 tools at 48% number is the one I keep coming back to, because it reframes tool overload as a measurable accuracy cliff rather than a vague smell. Splitting into special-purpose agents helps, but it moves the risk to the handoff, which is where I see most multi-agent systems actually break. I wrote up how I vet MCP servers before trusting them in an agent, the same trust concern from the tool side: kartiknvjk.hashnode.dev/how-i-evaluate-mcp-server… . Did accuracy recover once each sub-agent was back down to a handful of tools?