Lawrence Chen's AI setup
What I build
I work on cmux, a Ghostty-based macOS terminal built for running multiple AI coding agents side by side. I use a superrepo-style control repository above the source repositories to coordinate task topology, worktrees, skills, build rules, releases and verification.
Agents & harnesses
- Codex (current): Primary coding agent for focused tasks in isolated worktrees.
- Claude Code (current): Parallel coding agent in cmux workspaces; teammate mode can map to native cmux panes when useful.
- Built-in subagents (occasional): Bounded research or implementation delegation; no custom external harness.
- cmux (current): One task per workspace, with
Cmd+Shift+Ujumping to the newest unread agent notification for review.
Apps & terminal
- cmux (current): Terminal tabs, splits, browser panes, notifications and CLI/socket control.
- zsh (current): Default shell for scripts and development commands.
- GitHub (current): Source, issues, pull requests, checks and release coordination.
- Browser automation (current): Isolated browser sessions for web research and UI verification.
- Docker (available): Ephemeral container isolation for untrusted or network-facing inspection.
Machines & isolation
- Mac workstation (current): Runs cmux, agents and local review.
- Isolated worktrees (current): One worktree per change; shared primary checkouts remain untouched.
- Remote build and test machines (current): Builds and verification can run on isolated remote Macs or VMs when needed.
- Tagged dev instances (current): Unique app tags and sockets keep concurrent sessions separate.
Skills & instructions
- Repository instructions (current): Define worktree ownership, focused tests, builds, dogfood, release gates and secret handling.
- Reusable agent skills (current): cmux, Swift, browser, release, verification, performance, security and infrastructure workflows are selected per task.
- cmux primitives (current): I prefer composable terminals, browsers, notifications, workspaces, splits, tabs and CLI/socket automation over a fixed orchestration product.
How I work
- Start one task per cmux workspace and give the agent a focused brief.
- Let the agent read the relevant instructions, choose the needed topology and create isolated worktree(s).
- Start setup and build work early, then implement with focused tests and checks.
- Press
Cmd+Shift+Uto jump through finished-agent notifications, review the diff, build, artifact or bug fix, and reply to unblock the agent. - When the notification queue is clear, create another workspace and start the next task.
- Exercise the advertised path through the tagged app’s CLI/socket entrypoint and capture UI evidence before handoff.
- Dogfood the tagged build, address concrete review or CI failures, and merge only after required checks and approval.
- Keep secrets, credentials, private endpoints, raw configuration, private project details and device identifiers out of shared content.