10 coding agents running in parallel.
Zero idea which one finished.
That was the problem.
They weren't writing code by hand anymore
Everything ran through Claude Code and Codex in the terminal
10 panes. 10 agents. 10 copies of the repo.
And two problems appeared almost immediately.
They couldn't tell what each agent was working on.
They couldn't tell when an agent had finished.
So they built cmux.com:
Native macOS Ghostty-based terminal with vertical tabs, notifications, and built-in browser
Each workspace shows the task, branch, ports, and pull request - so you always know what every agent is doing.
When an agent finishes, it moves to the top with a blue notification ring.
Like iMessage...
But for coding agents.
Who's adopting it first?
Staff and senior engineers.
It makes sense.
They're already used to breaking large problems into smaller ones and delegating the work.
The difference now?
Instead of junior engineers, they're orchestrating dozens of coding agents in parallel.
Even Austin and his co-founder don't work the same way.
One prefers worktrees.
The other doesn't.
That's why cmux ships primitives, not prescriptions.
Developers build the workflow that fits them.
Because nobody has figured out the "right" way to work with coding agents yet.
And as those agents write more of the code...
The bottleneck won't be software.
It will be human attention.
00:00:50 — Pivoting from enterprise AI to dev tools
00:01:25 — Leaving the IDE for the terminal
00:02:00 — The problem that led to Arc-inspired workspaces
00:02:45 — Notifications for coding agents
00:03:25 — Why developers are moving to terminal-first workflows
00:04:30 — Why senior engineers adopt cmux first
00:05:50 — Primitives over prescribed workflows
00:07:00 — The future of autonomous code review
00:08:40 — Why human attention becomes the bottleneck
00:09:20 — Growing to 14.5K GitHub stars
00:11:15 — Building cmux for speed and performance