How autonomous is too autonomous

Autonomy in agents is a slider, not a switch, and the right setting depends on the task more than the technology. The instinct to push the slider toward “fully autonomous” is real because it makes the demo look magical. The cost shows up in the support queue, when an agent ran an irreversible action under uncertainty and nobody was there to catch it.

The settings worth distinguishing

At the assistive end, the agent suggests, the human commits. At the supervised end, the agent acts but every consequential action requires confirmation. At the bounded-autonomous end, the agent acts within a sandbox — limited tool surface, limited scope, fully reversible. At the fully autonomous end, the agent runs without supervision and the burden of reliability is entirely on the system. Most production agents should live at “bounded autonomous” and most teams accidentally drift past it.

The drift you don’t notice

The drift toward more autonomy happens one tool at a time. A new capability is added because a user asked for it; a confirmation step is removed because users complained about friction; a permission is broadened because edge cases didn’t fit the original boundary. Each change is reasonable; the cumulative drift produces an agent more autonomous than anyone designed.

The right level of autonomy is the one where your worst-case incident is still recoverable. If you cannot answer “what’s the worst-case incident,” your agent is probably more autonomous than it should be.