Terminal Agents Suffice for Enterprise Automation

A team of eight researchers including Patrice Bechard and Orlando Marquez Ayala has published a provocative pre-print paper challenging the complexity of current enterprise AI automation systems. Their research, submitted to COLM2026, argues that the growing interest in sophisticated agent architectures like Model Context Protocol (MCP) and web-based graphical interface agents may be over-engineered. Instead, they demonstrate that coding agents equipped with nothing more than terminal access and filesystem permissions can effectively solve enterprise automation tasks by interacting directly with platform APIs.

The team evaluated their hypothesis across diverse real-world systems, comparing terminal-based agents against more complex architectures. Their findings show that these simpler, low-level agents not only match but often outperform their more complex counterparts. The research suggests that strong foundation models combined with basic programmatic interfaces are sufficient for practical enterprise automation, potentially reducing both cost and operational overhead significantly. This challenges the current trajectory of increasingly complex agentic systems and offers a more streamlined approach to AI-powered enterprise workflows.