Source note

Workbench: A TUI for parallel coding agents

Coding AgentsTerminal UIMulti Agent EngineeringDeveloper ToolsHuman AI Interaction

Workbench is a full-screen terminal UI for running multiple coding-agent CLIs in parallel. The excerpt describes a developer tool with tmux-backed persistence, file viewers, shell panes, and git diff tracking, but it gives no benchmark evaluation.

  • Developers using coding agents often juggle separate terminals, editor tabs, file previews, and git diffs, which makes parallel agent work harder to track.
  • Agent sessions can be lost or disrupted when a UI restarts unless the process state lives outside the interface.
  • The problem matters because multi-agent coding workflows need clear session boundaries, fast context switching, and visible file changes.
  • Workbench runs each workspace as its own agent session with an independent tab strip for agent panes, shell terminals, and open files.
  • It uses a private tmux server so agent and terminal processes stay alive across relaunches and hot-reload restarts.
  • It supports pluggable coding-agent harnesses: Claude Code by default, plus Gemini, Goose, OpenCode, and Cursor via a command flag.
  • It adds a file Explorer, read-only file viewers, and a live git working-tree diff inside the same terminal UI.
  • The app is built with Bun, React 19, and Silvery, with PTY support from Bun >= 1.3.5.
  • The excerpt reports no quantitative benchmark results, user study, latency data, or coding-task success rates.
  • It claims support for 5 named coding-agent backends: Claude Code, Gemini, Goose, OpenCode, and Cursor.
  • It claims persistent agent and terminal sessions through tmux, using the socket path ~/.workbench/tmux-ui.sock.
  • It claims rich viewers for syntax-highlighted text, Markdown preview/source, images, PDFs, video playback, and Mermaid diagrams.
  • It includes regression tooling: bun run typecheck, bun test, bun run check, and a Playwright screenshot suite.