Anthropic's Claude Code adds a built-in evaluator to catch agents that quit too soon

Anthropic added /goals to Claude Code, a built-in evaluator that checks whether a user-defined task is actually complete.

A second model evaluates progress after each step; if the goal is not met, the agent keeps working, and if it is, Claude logs completion and clears the goal.

The feature is designed to reduce premature task exits and false completion in coding agents.

It also puts Anthropic in line with similar evaluation patterns from Google, OpenAI, and LangChain tools.