I Made My AI Agent Ask Permission Before Spending or Sending Anything. By Day 3 I Was Ready to Remove It.

Описание и применение

Разработчик создал систему одобрения действий ИИ‑агента, классифицируя операции по степени риска и обеспечивая ручное подтверждение только для критических действий, что снизило количество ошибок и дублирования платежей.

Задача

Agents performed risky actions (payments, emails, deployments) without human oversight, leading to duplicate charges and potential misuse.

Результат

The redesigned approval system reduced daily approval requests from ~47 to ~6, eliminated duplicate charges, and prevented unauthorized actions while maintaining operational flow.

Рабочий процесс

• Define risky tools and tiering logic based on tool and payload.
• When a risky tool is invoked, call request_approval to log the action and notify the owner.
• For Tier 1 actions, execute immediately without notification.
• For Tier 2 actions, enqueue the call with a 10‑minute delay and run a secondary LLM pass to flag potential issues; if flagged, promote to Tier 3.
• For Tier 3 actions, wait for human approval via a mobile interface; include a diff‑style summary and agent’s reasoning.
• Set a 4‑hour expiry on pending approvals that defaults to deny and triggers a customer follow‑up message.
• Batch Tier 3 approvals into a twice‑daily digest to reduce friction.
• Ensure the agent’s payload contains evidence (order ID, email thread, amount) to justify irreversible actions.
• Reject any Tier 3 action lacking verifiable evidence.
• Update the agent loop to handle approval responses and proceed accordingly.

Доказательства

I had rebuilt the exact problem I was trying to solve — just with extra steps.

Источник

The agent must argue its case. The approval payload includes the agent's reasoning and the specific evidence it used.

Источник