Anthropic has released a new feature called '/goals' for its Claude Code AI coding agent that addresses a common problem in enterprise AI systems: agents stopping work before tasks are actually complete. The feature formally separates task execution from task evaluation using two distinct AI models.

The new system works by having users define specific goal completion conditions through prompts, such as requiring all tests to pass and lint steps to be clean. Claude Code then operates in its normal loop of reading files, running commands, and editing code, but a separate evaluator model - Haiku by default - checks after every step whether the defined conditions have been met. If not, the agent continues working until the goals are achieved.

This approach differs from competitors' solutions to the same problem. OpenAI allows models to decide when tasks are complete but permits users to add their own evaluators. Google's Agent Development Kit and LangGraph require developers to manually configure evaluation logic and termination conditions. Anthropic's solution makes independent evaluation the default setting.

The issue of premature task completion has become a significant concern for enterprises deploying AI agents in production environments. Multiple sources report cases where coding agents finish their runs and show green pipelines, only for developers to discover days later that portions of code were never compiled or critical steps were skipped.

Security experts are also raising concerns about AI agent deployment more broadly. At the recent RSA Conference 2026, industry leaders highlighted authorization gaps where authenticated agents access data beyond their intended scope. Cisco's chief security officer reported that such incidents are occurring regularly across customer environments.

Meanwhile, other developments in the AI coding space include Microsoft reportedly canceling some Claude Code licenses after initially providing access to thousands of developers in December, and the release of open-source tools like Raindrop AI's Workshop for debugging AI agents locally.