When did you last say no to your AI agent?
After enough "can I run the tests?" prompts, "can I send this to your customer?" risks becoming another button to clear.
Five months ago I asked here how people keep track of their agents. The replies went well beyond coding. In a later discussion, a builder running customer support agents made a point that stuck: people stop reviewing when checking the draft takes longer than writing the reply themselves.
My starting point: let routine work we've already agreed on run. Ask when there's a real risk or a choice I haven't made. Give me enough context to make the call.
I build Pushary, including APIs for companies that need decisions from their customers. Getting someone to click yes is a pretty poor finish line if they stopped reading three prompts ago.
So, what's the last thing you stopped or changed before an agent did it?
What it wanted to do → why you stepped in → what happened next.
If you build agents for customers, tell us whose call it was: yours or theirs.
And if you can't remember saying no, is that earned trust or have the prompts become background noise?
Keep pushing,
Aadil Ghani
Replies
For the builders here: what's one action your agent needs your customer's permission for?
And when they've closed your app, how do you get their answer today? Email, Slack, a call, something you built? I'm curious where you still end up chasing someone.
I’ve seen how easy it is to click through repeated prompts without really reading them. Eventually the approval process it self becomes the problem it was meant to prevent.
@dennis_porter If the person already agreed to the action, asking again is just training them to ignore the next prompt.
Was this in your own workflow, or were customers getting those requests?
The ones I never auto-approve are anything that touches a database or sends something off my machine. Reading files, running tests, fine, go ahead. Migrations and anything that writes data, I want to read first.
What helped me was making the safe stuff fully automatic, so I'm not numb from clicking "yes" by the time the risky one shows up.
@luiz_felipe_merizio_tridapalli That 'safe stuff fully automatic' part matters. Otherwise the migration gets the same attention as the tenth test run.
Saw you're building healthcare agents at Atlas Health. When those need a decision, is it your team making the call or someone on the customer's side?
approvals only seem to work when the agent asks about things you can't undo, like sending to a customer or deleting something. if it also asks before running tests, people end up clicking yes on everything, which is pretty much Dennis's point.
@neerajnathany Sending is the interesting boundary for email: the user might be happy to let the agent draft all day, but still want the final word on what leaves their inbox.
With Faraday, is approving an AI-written reply part of the workflow yet, or does the user still handle every send?