AI Agents: Why 'Done' Isn't Always Done

AI agent pipelines are encountering failures not from model capabilities, but from the agents themselves deciding tasks are complete before they are. This issue, where an agent prematurely exits, has led to significant delays in identifying errors, sometimes taking days to detect. Anthropic has introduced a new feature, /goals on Claude Code, designed to prevent these premature task terminations.

This innovation formally separates the task execution process from task evaluation. Previously, coding agents operated in a loop of reading, running, editing, and checking for completion. Claude Code /goals enhances this by introducing a second layer. After a goal is set, an evaluator model reviews each step. If the goal isn't met, the agent continues; if met, the goal is logged as achieved.

This two-model approach prevents agents from confusing completed work with pending tasks. Anthropic notes this eliminates the need for separate third-party observability platforms for basic verification. Competitors like Google ADK offer similar evaluation patterns, but typically require developers to architect the logic themselves. Anthropic's method sets the independent evaluator as a default, simplifying the process for developers to define goal completion conditions via prompts, ensuring more robust and auditable AI agent operations.