Claude Code adds /goal run loop

This is a small command with a big product implication: Claude Code is moving from “help me code” toward “own this outcome.” The interesting part is the evaluator separation, which makes the loop feel more deliberate than a raw auto-run mode.

• –`/goal` starts immediately from a written completion condition, then re-checks after each turn
• –Anthropic says the evaluator is separate from the worker model, which should reduce self-confirmation bias
• –The feature makes Claude Code feel closer to a long-running agent workflow than a chatty coding assistant
• –It also raises the usual autonomy tradeoff: the better the goal spec, the better the run
• –For teams, this is most useful for test-fix loops, refactors, and other tasks with crisp exit criteria