Open Microsoft Word. Ask Copilot to help you write. Notice where it lives? ... a sidebar. A chat widget bolted onto the edge of an interface designed in the 1990s. Meanwhile, the ribbon toolbar with its nested menus and modal dialogs sits unchanged, untouched by the intelligence supposedly transforming knowledge work.

This isn't a design oversight. It's a structural inevitability.

The sidebar is the only place Copilot can live. Integrating intelligence into the ribbon would require rewriting the interaction logic of thirty years of legacy code. So we get a chatbot not because it's the best user experience, but because it's the only DOM element that accepts text-in and returns text-out without breaking the existing state machine.

I want to argue that this pattern reveals something deeper: Graphical User Interfaces—once the great liberators of human computing—have become technical debt for agentic systems. And the reason lies not in aesthetics or ergonomics, but in protocol theory.

The GUI as Contract

We typically think of interfaces as views ~ 'visual representations of underlying functionality'. But this framing obscures what GUIs actually are: communication protocols between human and machine.

Every button is a hard-coded promise. Every dropdown constrains the space of possible actions to its enumerated options. Every modal dialog enforces a particular sequence of decisions. The interface isn't describing capability; it's defining it. To change what the system can do, you must renegotiate the contract ~ which means rebuilding the UI.

This creates what I call back-pressure: the difficulty of updating the interface actively prevents the underlying intelligence from expanding. The GUI becomes a bottleneck not just for users, but for the system's own evolution, specifically its cognitive evolution.

Contrast this with text-based interfaces. A CLI is a stateless stream; declarative, open-ended, infinitely extensible. You don't need a visual widget to represent a new capability; you just need a new command, a new flag, a new argument. The protocol is loose enough to accommodate expansion without renegotiation. The trade-off is that text-based interfaces are harder to learn and use than GUIs. But, does this matter when you can just talk to an agent that is smart?

This is why coding agents are evolving so rapidly. Tools like Codex, Claude Code, Gemini-CLI, and others operate in terminal-adjacent environments where they don't need a "Refactor Button" to exist in some toolbar. They just need permission to write to the filesystem. The capability is the interface. Yes, there are tools like Google's Antigravity that is more graphical ~ but really they have a strong set of fixed panels & it is still very early days.

Conway's Law and the Frontend Bottleneck

The protocol rigidity of GUIs doesn't just constrain code. It constrains organisations.

Large GUI products require large frontend teams. These teams develop their own roadmaps, priorities, and backlogs. They become, in effect, gatekeepers. If the AI team wants to ship a new capability, they're blocked until the frontend team can build the UI to expose it. The intelligence is ready; the interface isn't.

This is Conway's Law playing out in real time: the structure of the software mirrors the structure of the organisation building it. Complex GUIs beget complex teams beget complex coordination overhead.

In TUI-first products, this friction largely disappears. Capability and interface collapse into the same artifact. There's no "AI team waiting on Frontend's Q3 roadmap." The engineer who builds the reasoning also ships the interaction. This organisational fluidity—not just technical simplicity—explains why agentic coding tools are outpacing enterprise AI features embedded in legacy applications.

The sidebar chatbot isn't just a UX compromise. It's an organisational white flag.

The Human Cost of Protocol Shift

Before this reads as a CLI manifesto, we need to reckon with what we lose.

GUIs solved a genuine problem: the terror of the blank prompt. We literally had to train people to use command based systems like DOS or the IBM Mainframes ~ but this was not a totally insurmountable problem. Humans prefer recognition over recall. We'd rather scan a menu for the option we need than remember the exact incantation to summon it. Buttons provide affordance & they imply their own function. A command line provides none. It demands that users already know what's possible.

This is the "Gulf of Evaluation" that HCI researchers have studied for decades. When you remove the GUI, you remove the user's ability to see the boundaries of possibility. You remove their sense of control. And for many users—including those with dyslexia, visual processing differences, or simply non-technical backgrounds—you remove access entirely.

There's also a subtler loss: professional identity. People define their expertise by their mastery of complex interfaces. "I'm an Excel wizard." "I know Photoshop inside out." Strip away the GUI, and you strip away the terrain on which that mastery was built. Resistance to AI tools isn't always fear of job loss; sometimes it's grief over the obsolescence of hard-won skills.

And then there's safety. GUIs act as speed bumps. Multi-step wizards prevent users from accidentally deleting production databases. Confirmation dialogs force a pause before irreversible actions. The "back-pressure" I've been criticising is also a safety valve. A TUI agent with filesystem access essentially has sudo over your work. Removing the friction of the UI also removes the constraints embedded in the UI's workflow.

The Verification Bottleneck

Here's what I think the current discourse misses: the frontier problem isn't input. LLMs have largely solved the input protocol. We can now express intent in natural language rather than learning arcane command syntax. The agent handles the recall; we supply the goal.

The frontier problem is verification.

In a GUI, state changes are visible. The file moves from Folder A to Folder B, and I watch it happen. In an agent workflow, magic happens. The agent ran seventeen commands, modified four files, and made three API calls. What actually changed? Text streams are notoriously difficult to audit. By the time I've parsed the terminal output, I've lost more time than the automation saved.

This is where I suspect the next wave of innovation will focus: not on better input modalities, but on better verification protocols. What does "visibility" look like for agentic actions? How does the user make sense? Perhaps structured logs with semantic diffs, perhaps a sensemaking apparatus that is yet to be invented (yes, this is on the top of my current research todo list). Perhaps agents that narrate their reasoning at adjustable granularity. Perhaps .. and this is speculative .. AI-generated summaries of AI actions, creating a verification layer as intelligent as the execution layer. We really do not know what works, what scales & interestingly, maybe different users need different sense making tools.

The teams that solve verification & sensemaking will unlock agentic adoption in high-stakes domains. The teams that don't will find their agents restricted to sandboxes where mistakes don't matter.

The Synthesis: Generative Interface

So where does this leave us? Stuck choosing between the rigidity of GUIs and the opacity of CLIs?

I don't think so. The synthesis is already emerging, and it goes by various names: Generative UI, Dynamic Interfaces, Ephemeral Applications. These are not new ideas, but previously the act of generating software was expensive & so, we had to be carefully about ROI & other such fun things.

The idea is simple: instead of choosing between a static dashboard and a text stream, the agent writes its own interface on demand. You ask about server load, and instead of returning a wall of text, the agent generates a React component, renders a graph, lets you interact with it, and then discards it when you move on (or writes a note to self that it can re-render if you ask again). The interface is as temporary as the question that prompted it.

This solves the core tension. You get the affordance of visual interfaces ~ something to see, something to click, without the rigidity of static layouts. You get the flexibility of natural language input without the opacity of pure text output. The protocol is loose on input, structured on output, and adaptive throughout.

We're already seeing early versions of this. Claude's artifacts. ChatGPT's code interpreter generating visualisations. Cursor's inline diffs that surface changes without drowning you in terminal logs. These are tentative steps toward a world where the interface is no longer a constraint on intelligence, but a product of it.

The Road Ahead

Let me hazard some predictions (in Jan 2026).

First, the gap between "agentic coding tools" and "enterprise AI features" will widen before it narrows. The former operate in protocol-flexible environments; the latter are trapped in legacy GUI contracts. Expect continued rapid evolution in the terminal, continued stagnation in the sidebar.

Second, verification infrastructure will become a category. Today's agents are trusted for low-stakes tasks and sandboxed for everything else. The bottleneck isn't capability; it's legibility. Whoever builds the "version control for agentic actions" ~ something that makes agent behaviour as auditable as a Git diff ~ unlocks the high-trust use cases.

Third, the accessibility question will become urgent. If the fast lane of AI-augmented work is purely text-based, we risk a new digital divide where only engineers wield the most powerful tools. Generative UI is one answer, but it's not sufficient. We'll need deliberate design effort to ensure that escaping the GUI prison doesn't mean abandoning the people GUIs were built to include, and there are many other accessibility considerations to address.

Finally, I suspect we'll look back on this period as the "Cambrian explosion" of interaction protocols. For decades, the GUI paradigm was stable enough that we forgot it was a choice. Now it's a constraint, and constraints breed alternatives. We're about to see more experimentation in human-computer interaction than we've seen since the Macintosh.

The GUI was a liberation technology in 1984. In 2026, it's an anchor. The interesting question isn't whether we'll cut the rope—we will—but what we'll find when we surface.

This essay is part of my ongoing exploration of AI-infused software systems. If you're building in this space, I'd love to hear what you're seeing.