Agent approval flows need clearer choices and safer defaults

Agent builders need better HITL when a person must approve or reject an ’s next action. The current approach does not easily support an approval time limit, a answer after time runs out, or a second person who can respond instead. People reviewing an approval request often cannot see what a choice will affect or how far a bad choice could spread.

They also need a way to ask the agent for more context before approving. A useful from Claude Code is automatic approval for actions that were already pre-approved and remembered. Real teams may also need to forward an approval request to a colleague when the first person cannot answer.

The open question is whether other ADK or SDK options support these HITL needs by , and which feature matters most.

Key points

  • HITL means a human checks or approves an ’s important action before it continues.
  • is described as missing approval expiry, replies, and backup responder support.
  • Reviewers need to understand the impact of each choice before approving.
  • Claude Code is cited as a useful example for remembering pre-approved actions.
  • Forwarding an approval request to a colleague is a practical need for team use.
Read original