Five things my 13 AI agents are never allowed to do

说明与用途

作者为13个AI代理实施了五个“门”系统，确保安全自治并通过单页合同和不可变日志实现完整审计。

问题

Managing autonomous AI agents risks uncontrolled actions that can incur costs, expose secrets, or cause irreversible damage.

结果

The gate framework prevented unauthorized actions, maintained auditability, and allowed agents to operate autonomously within safe boundaries, producing 1,235 items while keeping the human operator in control of critical decisions.

工作流

• Define five gates: money, auth, external sends, scheduled automation, destructive changes.
• Create one-page contract per agent specifying owner, mandate, data access, gates, escalation, kill rule.
• Maintain a single append-only ledger recording every item with status.
• When a gated action is needed, draft an envelope detailing the case, hard cap, stop-loss, and log tagging.
• Human reviews envelope, approves or rejects, and executes the action if approved.
• Monitor ledger for escalations and failures.
• Iteratively widen gates by adding envelopes as trust grows.

证据

People usually ask how I keep that from going wrong. The honest answer is boring: a short list of things no agent does on its own, and a long list of things they do without asking.

来源

The five gates: Money going live, Auth and secrets, External sends and posting, Scheduled automation, Destructive or irreversible change.

来源

Every agent has a one-page contract before it runs: owner, mandate, cadence, the data it may read and may never see, where it writes, which gates apply, who it escalates to, and a kill rule.

来源

One append-only ledger is the audit trail. One line per item: date | item ID | source | project | STATUS | note, where STATUS is BUILT, PARTIAL, ESCALATED or KILLED.

来源