Agent safety playbook: list every action, gate the irreversible ones behind human sign-off
rcleffke · reddit · 2026-11-19
A consultant shares a practical framework for governing AI agent permissions: before launch, enumerate every action the agent can take (including tools), mark the irreversible ones — refunds already issued, cancellations in effect, emails sent, payments submitted — and require human sign-off for those.
The five steps:
- Agents propose, not execute, anything irreversible
- Name a specific approver role with a backup and contact path
- Set review SLAs with a safe default: wait, don't proceed
- Log every irreversible action with who approved it
- Test the gate before customers meet the agent
He also lists five questions to ask whoever runs your agent; if you can't answer the first two, you don't yet know what the agent can do in your name.
Related event: Agents need human approval for irreversible actions(2 posts)→
More from coding & agent
- Netlify relaunches as AI app platform: describe-to-deploy, $130M ARR expected this year — thisiskp_ · 2026-11-19
- Factory expands Microsoft partnership: Droids on Azure, delegation via Teams — annbordetsky · 2026-11-19
- shadcn ships UI SDK for Solid 2.0 with agent-first docs and install prompts — shadcn · 2026-11-19
- Personal Agents Hackathon hits SF Nov 20 with $9,000+ in prizes, Meta and Cursor judges — code_star · 2026-11-19
- NVIDIA post-trained Nemotron with Factory entirely inside its air-gapped network — kuchaev · 2026-11-19
- Sentry's zeeg shares his LLM-powered email digest workflow to tame notification overload — zeeg · 2026-11-19