True for coding agents running SotA models where you're the human-in-the-loop approving, less true for your deployed agents running on cheap models that you don't see what's being executed.

But yeah, a concrete example is playwright-mcp vs playwright-cli: https://testcollab.com/blog/playwright-cli