Agents now ask before they take a risky action

Task Machine now checks what an agent is about to do, and holds risky actions in your Inbox until you decide. Posting a comment on GitHub, spending money outside Task Machine, changing production, or anything destructive waits for you, while everyday work such as editing files and running tests keeps going. We built it because agents often work while you are away, and you should hear about a risky action before it runs, not after.

Approve it once, or for good

The Inbox item shows exactly what the agent wants to do and why it needs you. Approve it once, or let that agent do the same kind of thing again for this task, for this chat, or always. When Task Machine can tell what the action touches, such as one repository, your approval covers only that. Reject it with an optional note and the agent tries another way or asks how to continue. The agent pauses while it waits and picks up where it stopped when you decide.

Some actions always wait for you

Deleting something large such as a home folder, wiping a production database, overwriting your main branch, or touching the files that control Task Machine itself always waits, even for an agent you trust. So does an action that follows instructions hidden in a web page or file. You can approve these only once, and an action that clearly breaks your company rules is blocked.

Ask less as you build trust

Every kind of action starts out asking first. As your decisions pile up, Task Machine can suggest letting a kind run on its own, and it still checks in now and then. You can take an approval back at any time in the agent's Autonomy settings.

Know the limits

The check is a model's judgment, so it can miss something, and permissions, credentials, and budgets remain the hard limits. Each check uses a little of your included usage, and workspaces that require EU processing skip it.

Read Inbox decisions for how held actions work and Autonomy for choosing what asks first.