Most agent demos end at the exciting part: the AI decides to do something, and it does it. That is a fine demo and a poor way to run your life. The moment an assistant can send email as you, move money or merge code, the question stops being “can it?” and becomes “should it, right now, without me?”
Fumu answers that question in one place, a single checkpoint that every action passes through.
One checkpoint, three answers
Before anything happens, the checkpoint looks at three things: how risky the action is, how much trust you have given that capability, and any rules you have written. It then gives one of three answers: run it, ask you, or refuse.
Because there is only one checkpoint, there is only one place to get safety right. A new capability cannot quietly skip it: everything Fumu can do goes through it, however it was added.
Trust is raised one capability at a time
Autonomy is not a single switch. Each capability has its own level:
- Observe: read and summarise only.
- Suggest: draft anything, but ask before any write.
- Act with notice: make reversible changes, tell you afterwards, and allow undo for a while.
- Autonomous within budget: act without asking, inside limits you set.
You might leave the Inbox agent at Suggest for months while the Research agent, which only ever reads, runs without asking from the start. When a capability has a clean record, you raise it. When it surprises you, you lower it.
Some rules never relax
A few actions always need a yes from you, whatever level you have set:
- spending money above a budget,
- messaging someone new as you,
- deleting data,
- changing credentials or permissions,
- installing a new capability.
These are the mistakes that are hard to undo, so they never become automatic.
What an approval looks like
When the checkpoint says ask, the task pauses and you get a card on Telegram or in the console. It shows what Fumu wants to do, why, and exactly what will change: the code, the email text, the amount. You can approve it, reject it or edit it. Edit matters more than it sounds. Most of the time a draft is nearly right, and fixing one sentence is faster than starting again.
Then the task waits. Nothing is lost while it does, even if something restarts in between. By default it waits 24 hours, reminds you once, and then cancels safely.
Approvals that teach
Every decision is recorded. When you approve the same pattern again and again (archiving newsletters, say), Fumu peeks over its shades with a suggestion: always allow this? Saying yes turns it into a written rule you can read, change and take back.
That is the whole idea. Fumu starts cautious, shows its work, and only gets more independent where you have watched it earn it.
