For years, tools like GitHub Copilot, Cursor, and Windsurf have dominated AI-assisted development. However, a quiet shift is redefining how these systems interact with software: instead of merely generating or editing code, they now operate directly on the system’s infrastructure. This pivot toward the terminal—a traditionally no-frills, command-line interface often associated with hackers and sysadmins—could revolutionize the future of AI-driven development.
The Terminal’s Comeback
The terminal, that stark, text-based interface, has long been overshadowed by sleek modern editors. Yet its power and flexibility make it the ideal environment for AI agents that go beyond writing code—compiling, configuring, and debugging entire systems.
Major AI players are already embracing this approach. Anthropic, DeepMind, and OpenAI have released CLI tools (like Claude Code, Gemini CLI, and CLI Codex), which have quickly become some of their most popular offerings. Though this shift has flown under the radar for many, it represents a fundamental evolution in how AI agents interact with machines.
*”Our bet is that 95% of LLM-computer interaction will happen through terminal-like interfaces,”* says Mike Merrill, co-creator of Terminal-Bench, the leading benchmark for evaluating these tools.
The Crisis of Traditional Tools and the Terminal’s Opportunity
While code-based solutions face uncertainty—like Windsurf, impacted by acquisitions and leadership changes—studies suggest their productivity benefits may be overstated. A METR analysis of Cursor Pro found that while developers believed they were 20-30% faster, they actually lost nearly 20% of their time due to the assistant’s suggestions.
This gap has fueled the rise of tools like Warp, the top performer in Terminal-Bench, which blends the best of an IDE and a terminal. Its founder, Zach Lloyd, argues that the terminal, operating at a lower level in the tech stack, allows AI agents to solve problems that would be impossible in conventional editors.
“The terminal is the most versatile place to run agents,” Lloyd asserts.
Terminal vs. Code: Two Different Approaches
The differences between these paradigms are clear when comparing their benchmarks:
- SWE-Bench (for code-based tools) focuses on fixing broken GitHub snippets.
- Terminal-Bench, however, evaluates holistic tasks—from setting up servers to compiling Linux kernels—demanding deep system understanding.
“What makes Terminal-Bench hard isn’t just the questions, but the complex environments agents must navigate,” explains Alex Shaw, its co-creator.
Even state-of-the-art models like Warp solve only about 50% of the challenges, highlighting the road ahead. But the potential is undeniable: these tools can already handle much of the time-consuming “non-coding” work developers face.
“Warp can autonomously set up projects, manage dependencies, and get everything running,” Lloyd notes. “And if it fails, it tells you why.”
Conclusion: A Terminal-Oriented Future
The shift toward the terminal isn’t just a trend—it’s a response to the limitations of traditional tools. As AI advances, its ability to operate in complex environments—not just lines of code—could redefine software development. Far from obsolete, the terminal is emerging as the next battleground for AI innovation.
By: Nestor Castillo, ForAllTechNews Director