we gave an agent our outreach and learned it cannot tell a draft from a send

we build agents for a living, so a while back we pointed one at the least glamorous thing we had, which is cold outreach to store owners.

the useful part worked. it could find owner names by reading the replies people leave underneath their own google reviews, which is a far better signal than anything you can buy in a list.

the part i had not modelled is that drafting and sending are the same shape of action to an agent. both are a tool call that returns success. one of them is recoverable in ten seconds. the other one spends a first impression, permanently, on a business we do not get to approach a second time.

i had been gating on whether a thing could be undone. that turns out to be the wrong axis entirely. it fires constantly on things nobody cares about, like writing a log line, and it stays quiet on the ones that actually hurt.

so the gate moved onto permanence instead. anything reversible inside a minute runs unattended. anything that spends a first impression waits for a person. and the gate does not ask the agent anything, because the agent cannot see how many businesses are on the list. it looks at the action type and the size of the audience, and both of those are things i can count without a model in the loop.

it cost almost nothing in throughput and it took most of the anxiety out of leaving the thing running overnight.