Human Approval

Autonomous agents at scale need an emergency brake. Human approval is the last line of defense when policies and guardrails aren't enough.

Human approval is two-factor authentication for agent actions — something the agent wants to do plus something a human confirms.

Scenario

Agent prepares $2,000 refund → approval card in Slack with customer context → manager clicks Approve → `issue_refund` executes.

Solution

In Security & Guardrails, apply Human Approval to this scenario: Agent prepares $2,000 refund → approval card in Slack with customer context → manager clicks Approve → `issue_refund` executes. Identify the inputs, run the technique, validate the output, and note one thing you would monitor in production.

Do this before moving to the next module — reading alone is not enough.

Open the Code Walkthrough below and run it locally. Change one parameter related to Human Approval (e.g. model, temperature, top_k, or tool name), observe the difference in output, and write 2–3 sentences explaining what changed.

Highlighted lines show where Human Approval happens in the code.

1async def execute_with_approval(action):  # key line for Human Approval2    if action.risk_score > 0.7:3        approval = await request_human_approval(action, channel="#finance-approvals")  # key line for Human Approval4        if not approval.granted:  # key line for Human Approval5            return "Action denied by approver."  # return the result6    return await action.execute()  # return the result

• Treating Human Approval as a black box without evaluation
• Ignoring cost and latency in production
• Skipping error handling for human approval

Quick recap — the most important points from this module.