·
5 min read
A confidence threshold is a product decision
Where you set the line between “draft and ask” and “act and log” says more about your company than any values page.
The dial everyone argues about
Every autonomous workflow in Vesper carries a threshold: the confidence at which the agent stops asking and starts doing. Set it too high and you have an expensive suggestion engine. Set it too low and you have an incident waiting for a date.
We’ve learned to treat that number as a product decision owned by the team closest to the consequences, not as a model parameter tuned by engineering. Finance sets the threshold for refunds. Security sets it for access changes. Each owner can see, in plain language, what the agent did at the current setting last week.
Making the dial legible
The threshold screen shows two lists side by side: actions the agent took on its own, and actions it would have taken at a slightly lower setting. Owners read both, and the conversation moves from abstract risk to concrete examples. That is usually when the number stops being controversial.