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

For many years, code editing tools such as Cardor, Windsurf and Github’s Copilot are standard for software supporting AI. As the agent of the AI departed more powerful and vibe coding, a subtle turn changed how the AI systems interacted with the program.

Instead of working on the code, they are increasingly built directly with the shell of the system. How is an important change in how the EU employed program development is a significant change – and despite the lower profile, it can be significant because the area is going from here.

The terminal is best known as the black-and-white screen you warns the hacker movies of ’90 years – Manipulation of many old schools and data of programs. It is not visually affecting as modern code editors, but if you know how to use it is an extremely powerful interface. Codic-based agents can write and debug code, terminal tools are often needed to get a program to something that can be used in written code.

The most clear sign of the transition to the terminal came from the main laboratories. From February, anthropic, Depremind and Openai, all team codes coding tools (claude code, twins Cli and Cli Codex and Cli Codex) and already are the most popular products of companies.

It has been easier to miss the sliding, because previously operated as previous coding tools under the same brand. However, under the hood, the agents had real changes in interaction with other computers of both online and offline and offline. Some believe that these changes are just starting.

“The combination of the leading terminal-oriented criterion through the interface of 95% of the LLM-computer interaction, 95% of the LLM-computer interaction Terminal-machine.

Terminal-based vehicles also enter themselves as the visible code-based tools start visible. The AI code editor is Duelinge with Windsurf, Chief Managers, splitted each other with dueling Employed by Google and the rest of the company obtained with cognition – to leave the long-term future of the consumer product indefinitely.

At the same time, new research offers programmers to overpring with ordinary means of productivity. A meter research Cursor Pro, Windsurf’s main opponent, found that they could perform more than 20% faster than 20% of the developers, the process was about 20% more slowly. In short, the code assistant actually cost the programmers.

He now left an opening for companies such as Warp, which is the best place in the terminal-machine. Warp, as an “Agentic Development Environment”, Ide Programs and Claude code itself as a medium foundation itself between the team linear means.

But Warp founder Zach Lloyd, this cursor, as a code editor for a code editor, we saw as a way to solve problems to solve problems.

“The terminal developer occupies a very low level in the stack, so agents are the most versatile place,” said Lloyd.

To understand how the new approach is different, it can be useful to look at the criteria used to measure them. Codes-based tools were aimed at solving Github’s problems, solving the basis of the SWE BENCH test. Every problem in SWE-Bench is an open issue from Github – in fact, a piece of code that does not work.

Models are repeated in the code until it finds something that solves the problem. The integrated products such as the cursor have built more complex approaches to the problem, but the Gitub / Swe-Bench model is still the basis of how these tools approach the problem.

Terminal-based means include a wider view, coding, coding, which is visible outside the code outside the code, but also more oriented assignments, as well as coding or removing problems.

In The problem of a terminal problemInstructions provide a decompression program and a target text file that complicates the agent for an engineer who compiles an interbiting algorithm. Another He asks the agent to want the source from the source, to want the source from source. Solving issues requires the type of solving the problem of programmers in need.

“TerminalBench is not the questions of how difficult it is, the agents say,” says Terminal-Bench Co-Creator Alex Shaw. “This is the environment where we placed.”

Kill this new approach means to solve a problem one step by step – the same skill that makes Agent AI so strong. However, even the most modern agent models are unable to handle all this environment. By solving only half of the problems, the terminal collected high scores – how difficult the benchmark and how much work should be done to unlock the full potential of the terminal.

Again, Lloyd believes that the terminal-based means is at a point where a developer can safely handle most of the coding work.

“If you do a daily basis to set a new project, identify addictions and get rid of working, you can do a lot in an autonomy to get a stick,” said Lloyd. “And if he can’t do that, you will tell you why.”