Agents Driving Terminals

With the Terminal Control ability, an agent can **open and operate interactive
terminal programs** — a coding assistant like Claude Code, a CLI, a REPL, anything. The
agent's tools are low-level and program-agnostic: open a program, read the screen, type
text or named keys, and wait for the program to react and settle. "Driving Claude Code"
is then just a recipe the agent follows, not a hard-wired feature.

Watch and take over. A program an agent opens shows up in your Your sessions
sidebar, badged AGENT with the command it launched. You can attach to watch it
live, and hit Pause to take over — while paused, the agent can't type, so you and
it never fight over the keyboard. Resume to hand it back.

Because this is effectively shell access, the ability is off by default and the
risky open/close actions can require your confirmation.

The mirror image — you driving a program through chat — is Terminal Tunnel (You Drive a Program). See also The In-App Terminal and Abilities.