Matched Capabilities, Different Behaviors: A Controlled Study of MCP and CLI Interfaces for Language Agents