Comment on: Openbase

This is the hard part and you're putting it front and center: which calls the agent makes alone vs. which ones it asks about out loud.We ship agents in a very different domain — they message customers and take payments for small businesses — and the rule we landed on after getting it wrong is: plan freely, execute visibly, but always ask before anything that touches money, anything that goes out to many people at once, and anything that changes configuration.The trap we hit: we tried letting the agent tune its own thresholds, with a guard scoring the outcomes. It optimized straight into the failure mode — a safe, wrong answer scores beautifully, because it generates no escalation and burns fewer tokens. We killed the autonomy and kept the human in the loop.Curious how you think about that when the approval is by voice, where "yes" is a very cheap thing to say.