Approvals and the kill switch

transcript

Delegation only works if you can stop it. Agent Hive gives you two controls for that: approval gates, which catch risky actions before they happen, and the kill switch, which halts work that is already in motion. Learn both and you can hand the company real autonomy without lying awake about it.

An approval gate is a rule that says a certain kind of action must come to you before it runs. Sending money, publishing something public, changing a customer record, deleting data: these are the defaults, and they are sensible. When an agent wants to do one of these things, the work pauses and an approval lands in your queue with the full context of what it wants to do and why. You approve, you reject, or you edit and approve.

The discipline here is to read the why, not just the what. An agent asking to send a $200 ad payment is not interesting on its own. An agent asking to send a $200 ad payment because last week's identical spend returned nothing is a decision you should probably reject. The approval shows you the reasoning trail so you can judge the decision, not just rubber-stamp the action.

You can tune the gates. If you trust the CEO with marketing spend under $50 but not over, set that threshold and the small payments stop interrupting you. The goal is to receive few approvals, each of which actually deserves a human. A queue full of trivial approvals means your gates are set too tight, and you will start approving without reading, which defeats the purpose.

The kill switch is the other control. It stops every agent in the colony immediately. Use it when something is clearly wrong: an agent stuck in a loop, a plan that turned out to be a mistake, a runaway you did not expect. Nothing is lost when you stop; work pauses and waits. You can inspect what happened, correct the course, and resume. The stop is a Chairman's privilege and it is always one click away.

There is a quieter version of the kill switch worth knowing: pausing a single agent. If one role is misbehaving but the rest of the org is fine, pause just that agent and let the CEO route around it. You rarely need to stop the whole company. But when you do, do not hesitate. A clean stop and a calm correction beats letting a bad plan run because you were not sure.

It is worth being concrete about which actions deserve a gate, because the instinct to gate everything is strong and wrong. Gate the actions that are hard to reverse or that cost real money or reputation: sending funds, publishing in public, changing or deleting customer data, signing anything. Do not gate the actions that are cheap and reversible: drafting, researching, summarizing, proposing. A draft an agent wrote can be thrown away for free; a payment it sent cannot. Set your gates to the line between reversible and not, and you will find the queue stays short and every item on it is real.

One subtlety: an approval is a teaching moment, not just a checkpoint. When you reject something, say why in a sentence. The colony remembers the reason, and the next time a similar action comes up, the reasoning behind it will already account for your past objection. Over a few weeks, a Chairman who explains rejections trains a CEO that proposes better actions and asks for approval less often. Silence teaches nothing.

Approvals and the stop are what make autonomy safe. They are not training wheels you remove later. The most experienced Chairmen keep tight gates on the few actions that can hurt them and wide latitude on everything else. That is not caution; that is good governance.

copy-paste prompts

Show me my current approval gates. I want to approve any outbound payment over $50, anything published publicly, and any change to a customer record. Everything under those thresholds, you handle and log. Update the gates and confirm what will still reach me.

Pause all work right now so I can review the current state. Summarize what each agent was doing when it stopped, then wait for my go before anything resumes.