Our AI teammates can read anything, but change nothing until I say yes
While filming a demo of NEXUPIRA this week, I kept coming back to the question every agent builder hits: how much should an AI teammate be allowed to do on its own?
Where we landed:
Reading is free. Teammates look things up themselves: the GitHub issue and the code, the Linear board, the latest messages in the customer's Telegram groups. They read before they answer.
Acting is not. A teammate never sends, comments or closes anything. What should happen goes under "Needs your OK" and becomes a confirm card for the owner.
One place to say yes. Everything lands in NEXUPIRA, so I review it there instead of jumping between three apps.
The result: the agents do the reading and the arguing in a meeting room, and the human only makes the calls.
Here is the 100-second version (demo project, sample data): https://x.com/JojoWu9627/status/2122192351884144857
How do you draw the line for your agents? Read-only by default, or let them act and undo later?
New sign-ups this month get a 7-day Pro trial: https://nexupira.com/signup?utm_source=producthunt&utm_medium=community&utm_campaign=nov-onehub
Next up: how our marketing and knowledge teammates plan, write and publish a real product launch together.


Replies
Read-only by default makes sense to me, with one design caveat: if every action needs a yes, people start approving without reading. I'd split it by reversibility. Things that are easy to undo (a draft, a label, a note) can happen on their own with a visible undo, and things that can't be taken back (sending, closing, publishing) go to the confirm card. It also helps if the card shows exactly what will happen, like the final message or the diff, rather than a summary, so the owner is making a real decision instead of clicking through.
The bit after the yes matters just as much as the yes itself. When ours sends an email for someone, the useful part is that it keeps an eye out for the reply and flags it when nothing comes back...