My AI Agent Did Something I Never Approved

An agent that acts without asking hands you a liability with a progress bar.

Most agent tools treat approval as an afterthought. The loop runs, and you learn what it decided from a summary, or from a customer replying to an email you never saw. Once an agent has posted, sent, or merged something without you, the damage is rarely catastrophic, but the trust is. You go back to doing everything yourself, and the tool joins the pile of things that almost worked.

Approval is the default

Task Machine starts from the other end. Every agent has an autonomy level, and the standard setting routes every consequential action, such as a send, a post, a merge, or a spend, to one inbox where it waits for your yes. Nothing user-facing ships until you approve it. The agent drafts and prepares the work, and you make the call.

That one change alters what agents are for. You can delegate freely, because the boundary is enforced by the system instead of by the model's judgment on a good day.

Turn trust up where it has been earned

Approval-first does not mean approving forever. As an agent proves itself on a kind of work, with clean drafts, passing checks, and fewer rejections, you raise its autonomy for that work alone. Routine actions then proceed on their own while risky ones still come to you. Budgets and verifier checks stay in force at every level, so even a fully trusted agent works inside hard limits.

Before risky work runs at all, the agent writes down what it plans to do and scores how far the change reaches, its blast radius, so surprising cases surface before anything executes.

How the first weeks feel

In the first week your inbox holds more decisions than you expected, because the system is showing you everything that used to happen out of sight. By the third week the routine work has earned autonomy, the inbox has thinned to real judgment calls, and nothing has gone out under your name that you did not see.