MCP vs CLI shifts agent behavior in 94.3% of paired runs

A new preprint reports that in 94.3% of 2,004 paired episodes, language agents given identical capabilities behaved differently depending on whether those capabilities were exposed through Model Context Protocol or a command-line interface. The study, 'Matched Capabilities, Different Behaviors,' posted to arXiv by Zhenning Shi, Liang Pang, and Yongqi Li, ran six models across three interactive environments using a framework the authors call MICA.

The authors are blunt about where the gap lands. 'Matched interfaces produce substantial differences in task performance and interaction cost,' they write, and the effects 'extend beyond surface validity to semantic action selection, failure pathways, and recovery.'

To test a fix, they propose DART, short for Decoupled Action Realization and Targeted Repair, which preserves what the agent decided to do while rewriting how it does it in whichever interface is in use. The abstract reports DART 'improves benchmark score in all six evaluated settings by up to 9.0 points and reduces MCP token cost in all three environments by up to 27.1k tokens per task.'

The abstract names neither the six models nor the three environments, and publishes no per-model or per-environment breakdown. Whether MCP or CLI tends to come out ahead on any specific task is not stated at this level.