# Approving actions

When an agent needs your go-ahead — to run a command, apply an edit, or answer a question — it pauses and asks right in the transcript. Nothing risky happens behind your back.

## Approving or declining a tool call

Under Ask permissions (and for risky actions under Accept edits), a tool call that needs sign-off shows an approval bar beneath its card — never hidden behind a chevron. Review what it’s about to do, then Approve to let it run or Decline to skip it. The turn continues either way.

## Answering an ask_user question

Sometimes the agent doesn’t need permission — it needs an answer. An ask_user question renders as its own interactive block in the transcript: type your reply (or pick an option) and the turn picks up with your input, no command line required.

## Autonomy sets the bar

How often you’re asked depends on the autonomy mode. Ask permissions gates everything; Accept edits gates only risky commands; Auto and Bypass let the agent proceed without stopping. Sub-agents in a swarm auto-approve under the session’s mode rather than asking separately.

> Bypass means no gateIn Bypass permissions there are no approval bars at all — the agent runs everything. Use it only when you trust the workflow end to end.