Command-line agents Updated 2026-09-06

Agent
control

Take hold of a running agent, start a fresh conversation, or direct its subagents. Across nine AI harnesses, the same command can mean different things.

You press Enter while the agent is working

Claude Codesteers the run
Codexsteers the run
DeepSeek Harnessqueues a follow-up
Hermesinterrupts the run
Pisteers the run
Grok Buildqueues a follow-up
Goosehas no busy prompt
OpenCodesteers the run
OpenClawsteers the run

Five of the nine steer. DeepSeek Harness and Grok Build queue separate follow-up turns by default; Hermes interrupts the active run. The Goose CLI only reads prompts between runs, so mid-run Enter is not a control. OpenClaw used to queue too, but changed its shipped default to steering in later releases. That is why this card exists.

A keyPress it. Nothing to type.
A UI controlClick it in the named graphical surface.
A commandType it, once, per message.
A modeSet it once. It stays set.
The model's own moveIt happens without you typing anything.
No native supportThe workaround is named in the cell.
Locally verifiedSeen in the installed build, not clearly documented.
Version driftNewer releases behave differently.

Select a control for instructions

When it hands the work to someone else

Eight of the nine ship subagents, and those eight spawn them the same way: the model decides. Pi deliberately leaves subagents to extensions. What separates the tools is what happens next — whether you can watch a child work, or only read what it chose to tell you when it was done.

Select a control for instructions