For years, code-editing instruments like Cursor, Windsurf, and GitHub’s Copilot have been the usual for AI-powered software program growth. However as agentic AI grows extra highly effective and vibe-coding takes off, a delicate shift has modified how AI programs are interacting with software program. As an alternative of engaged on code, they’re more and more interacting immediately with the shell of no matter system they’re put in in. It’s a big change in how AI-powered software program growth occurs – and regardless of the low profile, it may have important implications for the place the sphere goes from right here.
The terminal is greatest often called the black-and-white display screen you bear in mind from 90s hacker motion pictures – a really old-school method of operating packages and manipulating knowledge. It’s not as visually spectacular as modern code editors, however it’s an especially highly effective interface if you understand how to make use of it. And whereas code-based brokers can write and debug code, terminal instruments are sometimes wanted to get software program from written code to one thing that may really be used.
The clearest signal of the shift to the terminal has come from main labs. Since February, Anthropic, DeepMind and OpenAI have all launched command-line coding instruments (Claude Code, Gemini CLI, and CLI Codex respectively), they usually’re already among the many firms’ hottest merchandise. That shift has been straightforward to overlook, since they’re largely working underneath the identical branding as earlier coding instruments. However underneath the hood, there have been actual modifications in how brokers work together with different computer systems, each on-line and offline. Some imagine these modifications are simply getting began.
“Our large guess is that there’s a future during which 95% of LLM-computer interplay is thru a terminal-like interface,” says Alex Shaw, co-creator of the main terminal-focused benchmark TerminalBench.
Terminal-based instruments are additionally coming into their very own simply as distinguished code-based instruments are beginning to look shaky. The AI code editor Windsurf has been torn aside by dueling acquisitions, with senior executives employed away by Google and the remaining firm acquired by Cognition – leaving the patron product’s long-term future unsure.
On the similar time, new analysis suggests programmers could also be overestimating productiveness features from typical instruments. A METR research testing out Cursor Professional, Windsurf’s primary competitor, discovered that whereas builders estimated they might full duties 20-30 p.c quicker, the noticed course of was almost 20 p.c slower. In brief, the code assistant was really costing programmers time.
That has left a gap for firms like Warp, which at present holds the highest spot on TerminalBench. Warp payments itself as an “agentic growth atmosphere,” a center floor between IDE packages and command-line instruments like Claude Code. However Warp founder Zach Lloyd continues to be bullish on the terminal, seeing it as a solution to deal with issues that may be out of scope for a code editor like Cursor.
“The terminal occupies a really low stage within the developer stack, so it’s essentially the most versatile place to be operating brokers,” Lloyd says.
To know how the brand new strategy is totally different, it may be useful to have a look at the benchmarks used to measure them. The code-based technology of instruments was targeted on fixing GitHub points, the premise of the SWE-Bench check. Every drawback on SWE-Bench is an open challenge from GitHub — primarily, a chunk of code that doesn’t work. Fashions iterate on the code till they discover one thing that works, fixing the issue. Built-in merchandise like Cursor have constructed extra refined approaches to the issue, however the GitHub/SWE-Bench mannequin continues to be the core of how these instruments strategy the issue: beginning with damaged code and turning it into code that works.
Terminal-based instruments take a wider view, trying past the code to the entire atmosphere a program is operating in. That features coding but additionally extra DevOps-oriented duties like configuring a Git server or troubleshooting why a script received’t run. In a single TerminalBench drawback, the directions give a decompression program and a goal textual content file, difficult the agent to reverse-engineer an identical compression algorithm. One other asks the agent to construct the Linux kernel from supply, failing to say that the agent must obtain the supply code itself. Fixing the problems requires the type of bull-headed problem-solving potential that programmers want.
“What makes TerminalBench exhausting isn’t just the questions that we’re giving the brokers,” says Shaw, “it’s the environments that we’re putting them in.”
Crucially, this new strategy means tackling an issue step-by-step – the identical talent that makes agentic AI so highly effective. However even state-of-the-art agentic fashions can’t deal with all of these environments. Warp earned its excessive rating on TerminalBench by fixing simply over half of the issues – a mark of how difficult the benchmark is, but additionally how a lot work nonetheless must be performed to unlock the terminal’s full potential.
Nonetheless, Lloyd believes we’re already at a degree the place terminal-based instruments can reliably deal with a lot of a developer’s non-coding work – a worth proposition that’s exhausting to disregard.
“When you consider the day by day work of organising a brand new mission, determining the dependencies and getting it runnable, Warp can just about try this autonomously,” says Lloyd. “And if it may well’t do it, it should inform you why.”
