Auto mode is the permission tier that sits between approving every action by hand and `bypassPermissions`, which checks nothing. Until 2.1.158 it only ran against Anthropic's own API. Now it works with Amazon Bedrock, Google Vertex, and Foundry too, for Opus 4.7 and 4.8. Opt in with `CLAUDE_CODE_ENABLE_AUTO_MODE=1`. The mechanism hasn't changed: before each action a separate classifier reads the conversation and the pending command, then decides whether to let it through. Anthropic puts the false-positive rate at 0.4% on internal traffic, but the false-negative rate on a hand-built set of overeager actions is 17% — so treat this as convenience, not a safety boundary. Early reports also note Bedrock inference profiles don't always advertise auto mode even when it's configured correctly. If you run agent loops against a cloud provider, it's worth turning on; just don't point it at production credentials and walk away.