Opinion The command line interface is making a comeback because graphical user interfaces are a poor fit for autonomous agents, which could spell trouble for a lot of software – and software makers.
Forty years ago, all successful software boasted a well thought-out GUI and products without one had no chance. Good GUIs provided relief from the cognitive load associated with operating a piece of software, a burden that dropped precipitously as software converged on the same menu items and mouse gestures.
As the WIMPS (Windows, Icons, Menus, Pointers) interface became ubiquitous, learning to use one piece of software meant you knew almost everything needed to use every other. That’s now such a truism we don’t even think about it anymore, but was quite revolutionary back in the day.
It couldn’t last. The clarity of the classical GUI descended into Rococo flourishes of “ribbons” and “floating panels,” a parade of eye candy that brings a sugar crash of confusion. Today we’re likely to spend minutes hunting around for a menu option and asking ourselves “It was here last time I looked, did an update take it away?”
Developers were revising their reliable UIs, replacing them with something less useful; consider the notorious case of Apple wrecking its industry-leading FinalCut Pro (FCP), thereby handing the market to Adobe.
Change an interface too much, or too often, and all of that cognitive load returns. Software meant to be an invisible handmaiden to productivity becomes a burden: I quit FCP for Premiere and never looked back.
Autonomous agents are now living the same nightmare – as much as a machine can – as we instruct them to use computers on our behalf and they must try to learn about bad GUIs so they can drive them.
Agents must snapshot a screen, feed the image into a language model, analyze the results, change the approach they use to attempt executing our desires, and do that repeatedly until they finish whatever task a human set.
Bad interfaces make this process slow and needlessly complex: a UI that began as a way to ease the burden on humans has ended up being a deadweight for both humans and our agents.
Nowhere is this more visible than in the software that nearly everyone uses: Microsoft’s Office365Copilot (did I get that right? they’ve changed the name a lot). Three years ago Microsoft stuffed its Copilot AI software into every one of those apps, believing that after constantly adding apps to its suite without significant price rises it had finally delivered the next generation of office automation and found a way to meaningfully increase the cost of subscriptions.
Microsoft got it wrong. The software giant managed to be too early and in the wrong place, having misunderstood how people use AI. Redmond put their assistants inside the apps. No one really has any use for them there, where they lack anything except for local context. Cut off from the world, those assistants couldn’t do very much – and folks quickly learned to ignore them.
Instead, we see people using agents – such as Openclaw, Nanoclaw, and much else besides – to operate the apps. The agent sits outside the app, using that app as one of the many tools it has access to as it orchestrates a solution to a problem.
To do that, an agent needs a universal interface that’s easier to operate than the GUI.
So say hello – again – to the command line interface.
*nix neckbeards and PowerShell power users will claim the CLI never really went away – and they’re right, but they’re also a vanishingly small percentage of the computer-using population. For the rest of us, the CLI was something we might see only when things go very wrong with our computers.
Suddenly, the CLI is the best way for us to get work done with our computers. No one wants to spend the day wrestling with Outlook. We just want to say “schedule a meeting” to our agent. A CLI makes that sort of task easy for both agents and humans. As the GUI loses relevance for the next generation of agent interfaces translating the ambiguities of human language into actions, CLIs become the interface. Everything else is interference.
Google recognised this and leaned into it, last week releasing gws, a CLI for Google Workspace. The installation adds ‘skills’ to teach your agent how to fully manipulate all your Google Docs, Drive, Calendar, GMail, Sheets – the works. The whole universe of “office” software – previously thoroughly dependent upon a human sitting behind a desktop or mobile GUI – has become just another tool that an agent can operate to help you achieve your goals.
Microsoft might be in store for a penny-drop moment when company brass realise its flagship software has just taken a severe blow and been outclassed by a bit of interface glue.
Redmond won’t be the only victim. Nearly every major software provider will soon find itself scrambling to develop CLIs for their products, hoping to get to market before a competitor or hyperactive vibe coder provides exactly the same functionality through an agent-compatible interface. That’s what a ‘SaaSpocalypse’ looks like.
Case in point: my Openclaw agent Clawdine retrieved our full chat history from logfiles, exporting it into something I could use, because I could not work out how to export it from Telegram. (I suspect it’s a premium feature, sigh.) My agent enabled me to route around the damage from software grown baroque. It won’t be long until we’re all doing things like this, all the time. That will change software forever. ®