What changed
Across one six-operation repository task, seven agent scaffoldings and five models produced paired MCP-to-CLI cost ratios from 0.43x to 29x; scaffolding dominated and agents often ignored their assigned interface.
Why it matters
Model and interface labels alone cannot predict agent cost or reliability when the surrounding harness changes prompts, tool exposure, recovery, context, and actual execution behavior.
What to do
Benchmark the exact model-harness-tool combination, record which interface was really used, verify final repository state, and compare cost per completed task rather than tokens per attempted run.
Caveat
The study covers one private-repository task with six operations; only 13 strictly paired ratios were available, agents frequently violated interface assignment, and the author-run preprint does not show that CLI is generally cheaper than MCP.