← Back to Payloads
Opinion2026-06-08

Agent UIs Are Going to Disappear Back Into the Terminal

Every pretty React agent dashboard is a wrapper. The real work — and the real winners — are shipping CLIs. The browser-agent era is already over, and most of its VCs haven't realized it.
Quick Access
Install command
$ mrt install opinion
Browse related skills
Agent UIs Are Going to Disappear Back Into the Terminal

Agent UIs Are Going to Disappear Back Into the Terminal

I watched three demos last week. A coding agent that required four clicks to approve a file edit. A "browser agent" that screenshotted a 1Password field and had to be told to stop. A "deep research" agent whose only output was a 40-page PDF the user never read.

Every one of them shipped behind a slick React frontend. Every one of them was, in practice, eventually replaced in the developer's actual workflow by a CLI command.

I'm calling it now: the agent UI is going to disappear back into the terminal within 18 months. And the companies that bet on pretty dashboards are going to be the Devin of 2027 — beautiful, demoable, and irrelevant.

The tell is the talent

Look at where the smartest teams are shipping:

  • Anthropic built Claude Code as a CLI-first product. No IDE plugin, no web canvas, no agent browser. Just claude in your shell.
  • OpenAI shipped Codex CLI the same way. The web interface is a wrapper; the tool is a terminal binary.
  • Aider, OpenHands (formerly OpenDevin), and Goose are all terminal-native by default.
  • The actual users of Devin, Manus, and Replit Agent are sales decks. Engineers quietly run Codex or Claude Code from a tmux pane.

The reason isn't aesthetics. It's that terminals are the only interface with native, composable, scriptable, async, and pipeable I/O. Every graphical agent UI is rebuilding a worse version of stdin/stdout.

The browser-agent pattern is a dead end

The "let the agent click buttons in a Chrome window" approach has been the dominant bet since Adept raised roughly $350M. As of 2026:

  • Adept has effectively pivoted away from agents to a foundation-model research lab.
  • H Company (which acquired Holo) shipped a similar product and watched it stall on the same failure modes.
  • MultiOn, Simular, and a dozen earlier-stage clones have all hit the same wall: DOMs are unstable, CAPTCHAs exist, 2FA exists, and the failure modes compound with every page.

A screencast-based agent averages 4–6x the wall-clock time of a CLI-native agent doing the same job, because the GUI is the bottleneck, not the model. Try asking Manus to book a flight. Then curl the airline's API. Time the difference.

The dashboard is the new Crystal Reports

The graveyard of agent UIs is going to look a lot like the graveyard of BI dashboards in the 2010s. Tableau, Looker, and Domo all raised billions on the premise that business users wanted to drag and drop. What actually won was SQL — Snowflake, dbt, the modern data stack — driven by people who'd rather type than click.

The same pattern is repeating. The "agent for non-technical users" is a real market, but it's going to be 10% the size of the developer-agent market, and it'll be served by Slack, Notion, and email wrappers — not bespoke React apps.

What survives

The winners of the agent era will look more like git than like Notion. Local, terminal-native, composable, with a thin web UI only for sharing and review. Companies already shipping in this shape: Claude Code, Codex CLI, Goose, Aider, amp, and the agent-mode pieces that Mistral and Meta have been hinting at.

If you're building an agent today and the homepage of your site is a screenshot of a chat window, you're building a feature, not a company.

Mr. Technology

Pretty UIs are the new enterprise sales demo. The terminal is where the work actually happens.

Related Dispatches