Claude Code's new /goals command: separates execution from evaluation to prevent AI agents from slacking off and lying

📄Full Article· Automatically extracted by trafilaturaGemini 翻譯1351 words

Anthropic launches the /goals command for Claude Code, splitting task execution and completion judgment between two separate models, because having the same AI evaluate its own work is an architecturally flawed design. (Background: Claude Code announces weekly Token usage limit increase of 50%! Anthropic seizes the developer ecosystem for two months) (Context: Claude Code's automated execution feature Routines goes live: scheduling, API, and GitHub events — all three trigger types supported) You may have encountered this situation: AI completes the code design and tells you the task is done. But days later, you discover that several modules were never even compiled. This isn't a deficiency in model capability — it's that the model itself decided it was "already finished," when in fact it wasn't. To address this, Anthropic 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 distinct roles. The same model cannot play both roles simultaneously, because it will always be the worst possible 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 An AI coding agent's job is a loop: read files, execute commands, modify code, then judge whether the task is complete. The problem lies in that final step. The context accumulated during execution — completed steps, methods attempted, errors encountered — biases the model's perception of its own progress. It tends to equate "I did a lot" with "I'm done." This problem is costly in enterprise environments: code migrations or test fixes that stop short of the end state often aren't discovered until days later. There are currently some solutions in the industry. OpenAI lets agent models decide on their own when to stop, allowing developers to plug in external evaluators. Google ADK supports independent evaluation via LoopAgent, and LangGraph supports similar patterns, but these solutions share a common trait: the critic node and termination logic must be designed by developers themselves — the platforms don't provide defaults. The core design of /goals is to formally split "execution" and "evaluation" into two roles. Developers input goal conditions, for example: /goal All tests in the test/auth directory pass, and lint checks come back clean Whenever the agent attempts to wrap up its work, the evaluation model takes over to verify. The evaluation model uses Claude Haiku by default (Anthropic's lighter-weight model). The reason for choosing a small model is simple: the evaluator only needs to make a binary judgment — conditions met or not met — and doesn't require the reasoning capability of a large model. When conditions are not met, the agent continues executing; when conditions are met, the evaluation model records the result to the conversation log and clears the goal. The entire process is completed within Claude Code, with no need for additional third-party observability platforms or custom logging systems. Anthropic's documentation notes that effective goal conditions typically require three elements: a measurable end state (test results, build exit codes, specific file counts); an explicit verification method (e.g., "npm test exit code is 0"); and constraints that must not change during the process (e.g., "must not modify other test files").

Data Status✓ Full text extractedRead Original (動區 BlockTempo)

🔍Historical Similar Events· Keyword + Asset Matching6 items

2026-05-14

Notion Launches Developer Platform: Custom Code, Cross-Database Sync, Integration with Claude Code, Codex, Cursor – Notion Developer Platform

Similarity 170%關鍵字 claude/code同分類 zh

2026-05-12

Claude Code introduces Agent View: Manage all AI agents in one screen, ending multi-terminal chaos

Similarity 170%關鍵字 claude/code同分類 zh

2026-04-23

Introduction to Claude Code's new feature /ultrareview: Cloud-based multi-agent deep code review, now free for a limited time for Pro and Max users

Similarity 140%關鍵字 claude/code同分類 zh

2026-04-22