Claude Code finally got a real auto mode.
Claude Code already had modes with varying degrees of freedom. In practice the choice came down to two extremes: either confirm literally every action, or enable Bypass permissions and hope for the best. The middle mode — "Auto accept edits" — sounds like a compromise but in practice only accepts file edits while every bash command still waits for your OK. Autonomy was near zero.
The new auto mode works differently. A separate classifier system evaluates each agent action in real-time: safe ones pass without confirmation, risky ones are blocked. The agent can finally work on its own without bugging you every ten seconds, but won't delete half your repository.
This is exactly the principle Anthropic described yesterday in an excellent (read it) article on harness design: the more autonomous the agent, the more important a separate "observer" that checks its actions. Autonomy without control isn't a feature — it's a bug.