How do I stop an AI agent taking an action that cannot be undone?

Put sebbi.pro Point of No Return in front of the agent. Before any action the agent asks the gate. Reversible actions pass at once. An action that cannot be taken back, such as deleting data, paying or refunding money, sending a message, publishing, deploying or signing, passes only with a way back the account accepts: the fingerprint of a fresh backup, the exact reversing action, or a waiting window during which anyone responsible can cancel it. With no way back, or above the account's own limits, a named person approves or refuses on their phone, and no answer means no. Every answer, approval, refusal and cancellation is sealed into a tamper-evident chain and timestamped into Bitcoin, and an action reported as done without clearance is flagged on its record permanently. The guard is one Python file with no dependencies and stops the action if the gate cannot be reached; agents on MCP use the sebbi_no_return_check tool.

How it works

1. Get a key at https://sebbi.pro/point-of-no-return.
2. Download the one-file guard from https://sebbi.pro/point-of-no-return/sebbi_noreturn.py and set SEBBI_KEY.
3. Wrap each irreversible function with @no_return(agent=..., action=..., way_back=...), or POST /x/noreturn/check before the action.
4. Set your rules with POST /x/noreturn/policy: accepted ways back, shortest waiting window, the amount above which a person is always asked.
5. Read the per-agent safety report at GET /x/noreturn/report.