The worst way to learn that your agent made a mistake is from the customer it made the mistake for.
That is how I used to learn. A customer would text to ask whether the agent was supposed to do what it had just done. By then the email was out, and the customer had already paid for the mistake in time or embarrassment.
Guardian exists because of those texts. It is one of the trust tools in SAIN OS, and the reasoning behind the whole system is in Why SAIN is building the trust infrastructure for AI. This piece covers one practical problem: knowing what your agent is doing while it can still be stopped.
The 1 AM problem
An agent works while you sleep. That is the appeal of Super Intelligence (SI) that acts for you. A contractor CEO who travels wants follow-ups to go out while they are on a plane. A title lawyer with four inboxes wants the document chase to keep going after hours.
Now suppose one morning there is an email in the sent folder stamped 1:02 AM. It went to a client. You did not write it. Two questions follow. Why did the agent send it? And is it the only one?
With most agents you cannot answer either. You can read the chat history, if there is one. You can ask the agent to explain itself, and it will write a confident paragraph that may or may not describe what happened. Neither is a record, and neither would have stopped the email.
Here is the same night twice. Drag the clock.
A log explains the damage after the fact
The usual answer is logging. Keep everything, and look when something goes wrong.
Logging matters. Anchor'd keeps a tamper-evident audit trail of every action, approval, and draft, so the question "why" has an answer. But a record on its own works like an autopsy. It tells you what happened after it happened. At 1 AM nobody is reading logs.
Supervision is different. When a new associate does something unusual, a good partner does not wait for the monthly review. Someone notices, the work pauses, and a person with authority decides.
Agents need the same thing, and the supervisor cannot be a person watching a dashboard all night. That would defeat the purpose of having an agent.
What Guardian does
Guardian watches your agents and lets a named person stop them. A human always decides. Three steps, in this order:
- It flags. Unusual behavior is noticed when it happens, before a customer complains.
- A named person can stop it. You stop the agent, and the next email does not go out while you work out what happened with the first one.
- It waits. Guardian does not decide for you. Sometimes it will have stopped something that was fine. You say so, and the agent goes back to work. Sometimes it will have stopped something that would have cost you a client. Either way a person made the call, and the decision goes on the record.
The waiting is deliberate. An agent that resumes on its own after a pause has only been delayed.
Why the decision belongs to the owner
I could have built this as an internal tool, a way for us to catch problems before customers did. I built it for the owner instead. The agent acts under the owner's name. The liability is theirs. The decision about what is acceptable should be theirs too, along with the evidence.
No honest vendor can promise that an agent will never be wrong, any more than a firm can promise that about its people. The promise we make is narrower. When something is off, you hear about it from Guardian before you hear about it from a customer, and nothing else happens until you say so.
SAIN gives every agent ten simple rules, and the people who direct agents follow them too. Three of them carry this piece:
- Ask before you act when you cannot undo the action.
- Stop when an authorized person tells you to stop.
- Keep a record of what you do. Do not change or delete the record.
You can hold any vendor to those three, ours included.
Four questions to ask your vendor
If you use an agent today, or are about to, put these four questions to the vendor. Pick the answer you got, and see who is really supervising your agent at night.
Two answers end the conversation. "You can review the logs" means you are the supervisor, and you were asleep. "We decide when it resumes" means the agent does not fully work for you.
If you answer to a reviewer, ask for one more thing: a held action from a real week, with the decision and when it was made. A program office faces the same question in a different form, covered in Why AI in a program office needs a record.
Where to start
Start with one agent and one inbox. Write down what normal looks like for that agent: who it writes to, what it sends, and when it works. Then name the person who gets the call when it steps outside that, and the person with authority to let it continue. Often it is the same person. That short list is the start of your supervision plan, whatever tools you use.
If you want a second pair of eyes on that list, or want to see an agent held and released on a workflow like yours, talk to us.