AI coding tools are shifting to a surprising place: the terminal

In recent years, AI-driven coding tools have dominated the software development landscape, with names like Cursor, Windsurf, and GitHub’s Copilot leading the charge. However, a noteworthy evolution is taking place as these tools begin to shift their focus from traditional code manipulation to direct interactions with system terminals. This change, though subtle, promises to reshape the future of AI in software development significantly. Terminals, often depicted in classic hacker films as black-and-white screens, may seem outdated compared to modern code editors. Yet, they serve as powerful interfaces for those who know their way around them. While AI agents are adept at writing and debugging code, terminals play a crucial role in transitioning that code into functional software. The transition to terminal-focused tools has gained momentum, particularly among major AI labs. Since February, companies like Anthropic, DeepMind, and OpenAI have launched command-line coding tools—Claude Code, Gemini CLI, and CLI Codex, respectively. These products have quickly risen to prominence, hinting at a significant shift in how AI interacts with computing environments. Alex Shaw, co-creator of the TerminalBench benchmark, believes this is just the beginning. "We envision a future where 95% of interactions between large language models and computers occur through terminal-like interfaces," he states. This potential comes at a time when traditional code-focused tools, such as Windsurf, face uncertainty due to corporate acquisitions and executive departures. Research further complicates the conventional wisdom surrounding productivity gains from code editors. A study on Cursor Pro, Windsurf's main competitor, revealed that while developers felt they could work 20-30% faster, the actual data showed a nearly 20% decrease in productivity. Amidst these developments, companies like Warp are emerging with innovative solutions. Positioned as an “agentic development environment,” Warp aims to bridge the gap between integrated development environments (IDEs) and command-line tools. Its founder, Zach Lloyd, emphasizes the terminal's versatility, suggesting it can address challenges that standard code editors cannot. The difference in approach between terminal-based and code-based tools is evident in how they tackle problems. While tools like Cursor focus on fixing open issues from GitHub, terminal solutions take a broader perspective, addressing not just coding but also essential DevOps tasks, such as configuring servers and troubleshooting scripts. The challenges presented by benchmarks like TerminalBench require robust problem-solving skills. For instance, agents are tasked with complex directives, such as reverse-engineering compression algorithms or building the Linux kernel from source. These tasks demand a step-by-step approach, which aligns perfectly with the capabilities of agentic AI. While Warp has achieved a commendable score on TerminalBench by solving just over half of its problems, it underscores the benchmark's difficulty and the untapped potential of terminal-based tools. Lloyd asserts that these tools can effectively manage much of a developer's routine non-coding work, streamlining project setups and dependency management. "If it can’t solve a problem, it will explain why," he adds, highlighting the practical benefits of this new approach to AI in software development.