the future of agentic coding is bottlenecked by a tmux split-pane race condition

agent teams in Claude Code have a terminal multiplexing problem. when you spawn teammates, the system splits your current tmux pane instead of creating a new window. this causes a race condition -- keystrokes intended for one agent land in another, and layout corruption compounds as more agents spawn.

what happens: you start a team with 3 agents. Claude Code calls tmux split-pane for each. your original pane shrinks. input focus races between panes. you type a response and half the characters go to agent-2's pane. the layout eventually becomes unusable and you're left force-killing tmux sessions.

why it matters: agent teams are the most interesting capability in Claude Code right now. the TeamCreate / SendMessage / TaskList coordination protocol works. agents can decompose problems, claim tasks, send messages, and merge results. but the tmux layer underneath is a single point of failure. every team session i've run has eventually hit layout corruption or keystroke interleaving.

the fix seems clear: spawn each teammate in a new tmux window (or a detached session), not a split of the current pane. splits share a single terminal's input buffer. windows don't.

filed as github.com/anthropics/claude-code/issues/23615.

the irony: the agentic future isn't blocked on context windows, tool use, or planning ability. it's blocked on a 30-year-old terminal multiplexer's pane-splitting semantics.