▶ YouTube Summaries

• YT :: https://www.youtube.com/watch?v=dLhcLqoff6k
• Original title :: I'm done with terminals

A lifelong terminal user argues the terminal is no longer the right place to drive coding agents. Theo's trigger: he rebooted, worked eight hours, and only then noticed he had never opened a terminal. The case is about managing many parallel agents, images in prompts, mobile access, and durability - none of which terminals do well.

How he got here

• Antigravity's agent-manager view - one hotkey showing agent activity across project windows - was the first hint, even though he found the rest of the product poor.
• Before that he was pestering Cursor for one-click buttons to copy a worktree path so he could cd into it in his own terminal. His split was roughly 50/50: Cursor for prompting, terminal for diffs, git history, running code.
• The Codex desktop app was the tipping point: pasted screenshots, copyable text with intact formatting, non-laggy scrolling, and above all clicking between parallel tasks instead of memorizing tmux hotkeys.

Why terminals stop scaling

Terminal setups trade a bit of upfront configuration for low ongoing overhead - but that assumes stable work. With projects of wildly varying age and one to ten parallel tasks, the rigid pane layouts break down: three panes for three things, a flex pane, then a sixth task arrives and the mental model collapses. He had ~30 tmux terminals on one Linux box and repeatedly lost track of which session or worktree work was in, sometimes asking Claude Code to go find the Claude Code session he had been working in. sesh (mentioned as his default terminal tool, with SSH targets grouped in a folder) got closest.

Images were another fight: he invested real effort in making command-V image paste work through his terminal multiplexer over SSH. Roughly half his prompts include an image - screenshot the error, say "fix".

SSH is the real complaint

Powerful and 30 years durable, but fragile the moment the connection is imperfect. He hit sticky keys typing into Claude Code over SSH on his own LAN because the session bound to Wi-Fi rather than his 10 GbE link. Mobile is worse - he bought and supports Terus, still uses it, but does not trust real work on a phone terminal. mosh and SSH3 exist; he does not want to maintain them, especially for phone use.

Codex's remote control from the ChatGPT phone app showed him the value - start work on the desktop, finish it while running errands - but it was buggy, macOS-only, and Codex's macOS security-policy design spawns five or more processes per thread and subagent, wrecking performance.

The requirements list

Image pasting; all models at subscription rates (the ~$8-14k of inference for $200); good remote control; best-in-class Linux support (worktrees in 2 s rather than 30 s-2 min, and his 32-core Ryzen AI Max Framework Desktop barely reaching a third utilization); good mobile; low battery impact; and above all a good way to manage many things at once. Terminals deliver exactly one of these.

The open-source gap made it worse: good GUIs were locked to one provider (Claude app with a Claude sub, Codex app with a Codex sub), each vendor shipping new features only on its own preferred surface, so he found himself defaulting to OpenAI models simply because he preferred the app. Open Code avoids the lock-in but loses the subsidized official harnesses.

T3 Code

He pre-empts the self-promotion accusation: the tool exists because of the belief, not the reverse. Built with Julius after a failed native SwiftUI/AppKit attempt (scrollable text views were the blocker; Chrome does that well, so Electron won), architected over a websocket rather than Electron RPC specifically to make remote control easier.

Julius - the best terminal-hating developer Theo knows, who does all git through the VS Code extension - drove it to reliability: worktrees, Claude, Open Code, Grok, Cursor (integration still rough). It uses the official agent SDKs, not -p hacks, negotiated with the labs so subscription use stays above board; the Codex team has publicly cited it as inspiration.

The feature he says changed how he works: pick a machine, send a task, quit the app entirely, and the work continues - resumable from another computer, from app.t3.codes, or from the open-source iOS/Android apps. Connect via npx t3@nightly pair-tailscale (pairing code or URL, plus serve) or the simpler npx t3 connect, which creates a persistent background instance on Linux and macOS.

The personal payoff: he no longer schedules his day around keeping a laptop lid open. He used to code in Ubers as protected focus time, and long agent runs killed that - close the laptop, lose the agent. He reports going from three or four PRs a week to as many as 20 in a day on heavy days, attributing a significant part of that to leaving the terminal.

Wider argument

He frames closed tooling as a regression: the industry went from closed editors, to Atom and VS Code making customizable open tools the norm, back to closed forks and closed agent apps - ironic now that agents make customization trivial if you have source access. His goal is the best open-source surface for AI coding, and he does not believe that surface is a terminal.

He does not care which GUI you use - T3 Code, Codex, Cursor's Glass, Gene, Conductor, or your own - only that you try one once you have more than one or two things running. Closing analogy: we are still in the skeuomorphic era of AI dev tools, doing things the old way because the new way is too much at once. His parting jab: "the only way the terminal isn't holding you back would be if you're not building big enough."