I'm an old school user that finds the AI agent TUIs too magical, so I experimented with a new agent UX. I implement mu as a shell plugin (zsh and fish for now). Pressing Tab enters agent mode, where 1. the command line prompt is changed to displays model name and context usage, and 2. a preexec hook sends commands to an agent instead of interpreting them as shell script.

Thats it. Each LLM prompt starts a regular process that reads stdin and writes stdout, with no terminal magics. The actually executed commands appear literally in zsh_history, just like other shell commands. This allows a seamless switch between normal shell work and agentic work, with mixed history that uses regular scrollback buffer.

I also experimented with a minimum agent design, where only a single bash tool is provided. Special commands are provided for reliable file editing and multi-modal input. So far it worked well for me, as a day to day command line helper.