Claude Code new release /goals command: separate execution and evaluation to prevent AI agents from slacking off and lying

Anthropic introduces the /goals command for Claude Code, separating task execution and completion judgment into two independent models, because having the same AI judge its own work is inherently flawed.
(Background: Claude Code announces a 50% weekly token usage increase! For two months, Anthropic is capturing developer ecosystems.)
(Additional context: Claude Code’s automatic routines feature is live: supporting scheduling, API, and GitHub event triggers.)

You might have encountered this situation: the AI completes the code design and responds that the task is done. But days later, you discover that several modules haven’t been compiled at all. This isn’t due to model capability issues, but because the model decided it was “finished,” even though it wasn’t.

To improve this, Anthropic has launched the /goals command for Claude Code this week. The logic is straightforward: the model executing the task and the model judging whether the task is complete must be two separate roles. The same model cannot play both, because it would always be the worst judge of its own work.

How do you keep Claude working until the job is done? Claude Code helps with this in a few ways, including one we shipped recently: /goal. pic.twitter.com/QtVPmwoKct
— ClaudeDevs (@ClaudeDevs) May 13, 2026

Why do AI agents “clock out early”

The work of an AI coding agent is a loop: read files, execute commands, modify code, then determine if the task is complete. The problem lies in this final step.

The accumulated context during execution—completed steps, tried methods, mistakes made—causes the model to develop biases about its own progress. It tends to equate “I did a lot” with “I am finished.” This issue is costly in enterprise environments: if code migration or testing stops before reaching the final state, it often takes days to discover.

Currently, some solutions exist in the industry. OpenAI allows the agent model to decide when to stop itself, enabling developers to connect external evaluators. Google ADK supports independent evaluation via LoopAgent, and LangGraph offers a similar mode, but these solutions share a common point: critic nodes and termination logic need to be designed by developers; platforms do not provide defaults.

One command, two models

The core design of /goals is to formally split “execution” and “evaluation” into two roles. Developers input goal conditions, for example:

/goal test/auth All tests in the directory pass, and lint checks are clean

Whenever the agent attempts to finish, the evaluation model takes over to verify. The evaluation model defaults to Claude Haiku (a lighter model under Anthropic). The reason for choosing a small model is simple: the evaluator only needs to make a binary judgment—conditions met or not—without requiring large model reasoning capabilities.

If conditions are not met, the agent continues executing; if conditions are satisfied, the evaluation model records the result in the conversation log and clears the goal. The entire process is handled internally within Claude Code, without needing additional third-party observability platforms or custom logging systems.

Anthropic states that effective goal conditions usually require three elements: measurable end states (test results, build exit codes, specific file counts); clear verification methods (e.g., “npm test exit code is 0”); and restrictions that cannot be changed during the process (e.g., “do not modify other test files”).