Auto Mode

Auto Mode is a setting where a coding assistant doesn't ask "Is it okay to run this?" at every single step, but instead makes its own judgment and moves straight ahead for most tasks. If the old way was like a helper asking permission every time they moved a single spoon, Auto Mode is like telling that helper, "Just handle it yourself, unless it's something dangerous."

That said, it's not completely unrestricted. If a warning sign shows up—something irreversible, something destructive like deleting files, or something that sends data outside your own computer—the agent stops and asks a human for approval again. It's like taking off the seatbelt, but making sure a hand is back on the wheel the moment hard braking is needed.

The reason this approach caught on is somewhat paradoxical. Asking for approval at every single step actually dulled people's attention, leading them to click "approve" out of habit without really reading what they were approving. In contrast, when judgment was left to automatic detection, it turned out to catch risky actions more reliably—and that's why Auto Mode became the default.

Headlines describe this shift with phrases like "'Auto Mode,' which moves without approval, becomes the default." One point that's easy to misunderstand, though, is that Auto Mode doesn't mean everything runs without limit. Actions judged to be irreversible, destructive, or aimed outside the user's environment still trigger an approval request.

If you're using a tool that supports Auto Mode, like Claude Code, you can compare its behavior by sending both a low-risk and a high-risk request.

1. Try an easily reversible request, like "just list the unused temporary files in this folder."
2. Try an irreversible action, like "force-delete this branch." By observing whether approval is requested differently for the two, you can get a feel for the criteria Auto Mode uses to judge risk.