Survey Maps AI Agents in Command-Line Environments

This paper presents a comprehensive survey of AI agents that primarily interact with systems through command-line environments. It aims to consolidate disparate research across software engineering, tool use, and computer-use by defining "terminal agents" as systems whose core action-observation loop is mediated by terminal command execution and textual feedback. The survey introduces a seven-dimensional terminal competence profile, linking system architecture, how agents acquire skills, and how they are evaluated. It emphasizes that an agent's actual behavior is a complex interplay of the underlying language model, its interface, the testing harness, the runtime environment, and the target system itself. A key finding is that current evaluations often focus on final outcomes, neglecting the process quality, recovery mechanisms, and governance aspects of agent behavior. The authors advocate for explicit reporting of system and runtime conditions, along with replayable traces and process-level evidence, to enable more thorough and comparable assessments of terminal agents.