p50 latency executor adds to tool calls is now down to ~45 ms still above my goal of 10ms, but i have some future plans of how this'll actually end up being faster than calling an MCP directly from your harness with batch endpoints and dynamic placement