tmux keeps coding agents reachable, but it cannot tell you they are done

The case for tmux in AI coding is practical: keep the process reachable, name the work and script the terminal. A crystl guide circulating on Hacker News is prompting developers to revisit that setup, while also asking where a terminal multiplexer stops being enough for several agents needing attention.

The discussion began October 19 and showed 153 points and 116 comments when reviewed October 20. That establishes current interest in an existing workflow, not a tmux launch or evidence that one setup has displaced the alternatives. Replies include recommendations for remote sessions, preferences for other terminal tools and complaints about the maintenance burden of managing many agents.

A small pair of primitives does useful work

The guide’s central combination is send-keys for input and capture-pane for reading a terminal’s rendered contents. A script can address a named session and window, send a prompt and inspect the result without a separate integration API.

Its details matter. Literal input uses send-keys -l, with Enter sent separately, so words that resemble key names are not interpreted as keypresses. The guide recommends capture-pane -J when joined wrapped lines are useful. These are workflow recommendations from the author, not an AITrending-tested orchestration package.

For concurrent repository work, separate terminals and separate working trees solve different problems. Named windows organize processes; a Git worktree isolates a checkout. The guide describes Claude Code’s --worktree --tmux path, notes that the tmux option requires a worktree, and recommends checking the installed CLI’s help. It does not establish a minimum Claude Code version. Agent installation, account access and provider charges remain separate from tmux.

A quiet terminal is not a completion event

tmux can keep a server-side agent running when its SSH client disconnects. It does not keep a process alive through a host reboot or a stopped tmux server, and it cannot make a sleeping machine execute work. A durable transcript needs its own recording mechanism.

The harder limit is meaning. Capturing a pane returns a rendered frame, not a structured report of tool success or approval state. A silence alert can mean an agent finished, is thinking, is blocked or simply stopped producing output. Matching terminal text for status is sensitive to changes in the agent’s interface.

The guide is published by crystl, which promotes its own macOS agent terminal in the closing section. Its product comparisons are vendor claims, not independent proof of superiority. The useful recommendation stands on a narrower basis: tmux is worth considering for detachable remote sessions and scriptable terminal access, while completion, approval and history requirements need explicit handling beyond a quiet pane.