I wanted to start a broader mistakes and lessons thread around MCP.
A useful angle is evaluation, but it also connects with performance. Different tools and workflows can produce very different results even when the goal is similar.
It would be interesting to compare practical examples...