In auto mode, policy runs first, then low-risk misses can be reviewed by a model, and anything uncertain falls back to human approval. OpenClaw says the reviewer model is separate from the main agent model, and approval prompts can still resolve to `allow-once`, `allow-always`, or `deny`.
This is a practical attempt to reduce approval fatigue without giving up operator control. It also makes OpenClaw more usable for enterprise-style environments where routine automation should keep moving, but privileged actions still need a second look.
On a gateway-host setup, OpenClaw documents `openclaw config set tools.exec.host gateway` followed by `openclaw config set tools.exec.mode auto` to enable the mode. If you want stronger review than the main agent, set a dedicated reviewer model such as `openai/gpt-5.5`.
Read Original Post →