Language agents are increasingly operating through external tools and interactive environments, making the interface between the model and its executable capabilities a core part of the decision loop. Model Context Protocol (MCP) and command‑line interfaces (CLI) can expose identical semantic operations while differing dramatically in the representation presented to the model. We introduce Matched Interface Comparison and Analysis (MICA), which derives both MCP and CLI from a shared semantic action contract and routes decoded actions through a common runtime. Across six models and three interactive environments, matched interfaces yield substantial differences in task success and interaction cost, affecting semantic action selection, failure pathways, and recovery strategies. Matched‑state probes pinpoint sensitivity to how capabilities are presented, realized, and fed back. Paired natural trajectories show structured divergence in 94.3% of 2,004 episodes, with early differences propagating throughout the interaction. Building on these findings, we propose Decoupled Action Realization and Targeted Repair (DART), which preserves validated semantic commitments while repairing interface‑specific realizations. DART improves benchmark scores by up to 9.0 points in all six settings and reduces MCP token usage by up to 27.1k tokens per task across the three environments. Together, the results establish interface representation as a direct intervention boundary within the closed‑loop decision environment shaping agent behavior.
Review