From One Tab to 16 Agents: DHH's Parallel Programming Setup in the Terminal

DHH says he spent nearly two decades in TextMate, starting around 2005, and even helped ship its first version. He was never shopping for a replacement; what pushed him out was necessity rather than curiosity: the move to Linux. That break became the first step toward today's agent-driven setup.

The conversation opens with a naming debate: should the new practice be called 'agentic engineering' or plain 'programming'? DHH makes no secret of disliking the former, and he and Lex settle on the latter. But the plain name carries a sharp thesis: working with agents demands a different toolbox.

In the old setup, DHH would sink into one problem and follow it through to the end; that deep immersion was his doorway into flow. An hour might yield a single lovingly worked file at twenty or thirty lines. Care, not speed, was the metric.

Agents break that rhythm because they are too fast and too slow at once: unlike a keyboard, they never answer instantly and need time to cook. Waiting on a single agent leaves you feeling idle and oddly useless, even when the agent itself is productive. DHH admits the feeling openly.

His fix borrows from the scaling lawbook of AI itself: throw more resources at a hard problem. With a handful of agents running in parallel, the picture flips; there is always a decision to make, a stuck agent to unblock, a fresh task to start. Flow returns, born of motion rather than depth.

The setup began with tmux panes, a metaphor everyone knows from tabbed terminals: open many tabs, put one job in each. The striking choice is what came next, or rather what did not: no move to cloud-hosted coding apps. DHH stays in the terminal and its text interfaces, arguing the modern terminal is simply a beautiful place to work.

Past a single machine, tmux stops being enough and Herdr takes over: a multiplexer that chimes when an agent finishes or needs a decision, showing each pane's state as idle, working, blocked or done. The video's auto-generated captions spell it 'Herder'; the actual name is Herdr. It now ships by default with Omarchy's fourth major release, and the project's own blog reports downloads past 700,000.

The hardware layer grew on the same logic: DHH bought GL.iNet Comet GL-RM1 boxes, small KVM units that take HDMI and USB and expose a remote computer through one web page and one password. Four retired mini PCs from a closet joined the network that way, reaching about sixteen threads across four or five machines, three agents each. He adds that faster agents mean fewer threads he can supervise.

The second enabler is a WireGuard-based private mesh over Tailscale: every computer in the Malibu and Copenhagen offices appears as part of one local network, wherever he happens to be. No firewall holes, no elaborate VPN setup; even his phone reaches the machines directly. The friction of bringing new compute online drops toward zero.

The payoff, by his own measure, is a leap: from one lovingly built file per hour to hundreds of generated lines. He stresses that line counts are a silly metric, yet the volume jump is undeniable. Neovim is no longer where code gets written but a project browser for skimming the changelog; a diff viewer called Hunk renders changes prettily yet hides the surrounding context he needs.

Underneath sits an operating-system thesis: agents love the Unix philosophy, because on Linux everything is either a config file or a command-line tool. What made Linux unpopular five minutes ago is now its biggest selling point. DHH proves it with a weekend stranded on a Mac Mini: praise for Homebrew aside, tools like Raycast expose no config file, key bindings demand mouse-click caveman setup, nothing automates. Lex's WSL story lands in the same place: a sandboxed Linux cannot let agents roam the whole machine.