The part of AI agents that worried us most wasn't whether they could do the work. It was whether we'd trust them to do it without watching every step. So three of the things we built into Bees are about that: agents that stop and ask, agents that keep what you told them, and a way to get more than one view before a decision.
Bees is a free, open source desktop app that runs a team of AI agents on your own computer. It went public today. Here's how those three pieces work, each with a short video.
1. Agents ask first
Finished work doesn't just go out. It waits under Needs your attention until you approve it, or send it back with a reason. You can also mark a stage as needing approval, and then its agent can't send, post, delete or pay through a tool until you say yes. When an agent isn't sure what you want, it asks instead of guessing.
Approval is off by default and you turn it on per stage, so you only pay the waiting cost where a mistake would actually hurt.
2. Tell it once
When a report is wrong, reject it and say why in plain words. You pick whether the note is for this run only or for every future run. If it's for every run, Bees writes your note into that job's playbook, so next time it gets it right.
The playbook is plain text you can read. You can edit or undo any change Bees made to it, which matters, because a rule you can't see is a rule you can't fix.
3. Roundtable: three views, one answer
Some questions don't have one obvious answer. Ask one in Roundtable and three agents debate it, each with its own job: one pushes for growth, one guards the money, one speaks for the customer. Then a lead agent writes one clear call, with every reason written down, so you can see what it weighed and disagree with it.
Where it runs
All of this runs on your machine. You connect your own AI: Codex or Claude Code from the command line, a hosted provider, or a local model. Your files, browser sessions and passwords stay on your computer.
- Download: bees.bot
- Code: github.com/Bees-bot/bees-desktop (MIT or Apache 2.0)
If you've run agents on real work, where did you have to add a human check? I'm curious whether it's the same places we found.
Top comments (0)