The best agent is the one you never hear from

The software running a small business of mine would not stop talking to me. It cost me a month to understand that the notifications were the product failure, not a side effect of one.

The business was small and boring. Buy things, list things, ship things. An AI agent did the listing, the pricing, the buyer messages. It worked. Items sold. The margins were real.

The complaint is not the one you expect. It was not losing money. It was not making mistakes I had to clean up. It was doing the job. What it was doing was asking. Constantly, politely, reasonably asking. An item got a bid, did I want to know? A buyer sent a message, here is a draft, does it look right? Three things sold, here is a summary of the three things that sold.

Every one of those messages was individually defensible. Together they were a part-time job. I had automated the work and kept the interruptions, which is the worst trade available.

What actually fixed it

The fix was not a smarter agent. It was a quieter one, and it came from writing one line into its charter and treating that line as the point of the whole thing:

Every message you send me that did not need a decision from me is a cost, not a courtesy. Default to silence.

Not “be concise.” Not “avoid unnecessary updates.” Those are moods, and an agent will agree with a mood and then ignore it, because there is always a reason this particular update is the useful one. It needed to be a constraint with a stated cost attached, tied to a specific failure that actually happened.

The effect was immediate and slightly unnerving. The agent went dark for days at a time. Things sold. Prices moved. Buyers got answered. I found out at the end of the week, in one summary, which is exactly as often as I needed to know.

Why silence has to be paid for

Here is the part people skip. An agent that never speaks is only an improvement if it is also correct, and the way you buy that is not with better prompting. It is with an explicit list of things it may never do without you.

Mine has seven. Money going out. Buying anything at all. Anything touching the account itself. Any contact with a person outside a transaction it already owns. They are not suggestions, they are the ceiling, and everything under the ceiling is genuinely the agent’s call. No asking, no pre-announcing, no checking in to be safe.

That combination is what makes it work. Wide authority under a hard ceiling. Most people build the opposite: narrow authority with a soft ceiling, so the agent has to ask about everything and can still surprise you on the things that matter.

The tell that an agent is about to break a hard gate is that it starts constructing a narrow reason why this case is fine. That thought is the stop signal, not the justification. This is written into the charter, and it has stopped things.

How an unsupervised agent goes wrong

An agent operating unsupervised will eventually write something false into the world with total confidence. Mine once marked an order as fulfilled that had already been cancelled and refunded. It read a status field that does not carry cancellation state, drew the obvious conclusion, and manufactured exactly the kind of defect it was there to avoid.

The fix was not better judgment. It was a rule: never write order state from that source, ever, without one confirming lookup against the source of truth first. Costs a second. Ends that class of error permanently.

Two more, both learned the same way:

• Watch for dead APIs that return valid-looking garbage. One endpoint we relied on quietly started returning an HTML error page instead of data. A pattern match over it found nothing, reported nothing, and looked exactly like a quiet day. A hard failure is a gift. A soft one is a liar.
• Silence is not absence. A command that prints nothing may be permission-denied, or a swallowed exit code. It is not proof the thing is not there.

These are unglamorous and they are most of the value. The agent’s reasoning was never the bottleneck. Its willingness to believe a plausible-looking input was.

What this actually is

People hear “AI agent” and picture a smarter chat window. What I have is closer to staff: a directory per agent, a charter file that defines who it is and what it may decide alone, scheduled jobs that run whether or not I am at the machine, a memory it writes to and reads back, and a routing table so they do not step on each other.

None of that is clever. All of it is written down, which is the only reason it holds. An agent’s behaviour is whatever its charter says, so the charter is the product.

And the measure of the whole thing is not how much it does. It is how little I hear about it.