Claude Code 四类循环的触发与停止边界

🚨 Anthropic's Claude Code team just laid out the 4 loops you can actually build with Claude, and each one hands off a different piece of the decision-making to the agent. "Loop engineering" gets used loosely enough right now that it's worth being precise about what actually changes between these four setups: what starts the cycle, and what decides when it stops. Here's the breakdown: 🔹 Turn-based: you prompt, Claude acts and checks its own work, then hands the result back. This is the loop running under almost every message you send, whether you've named it or not. 🔹 Goal-based (/goal): you define a measurable finish line up front. An evaluator model checks your condition every time Claude tries to stop, and sends it back to work if the bar hasn't been cleared. 🔹 Time-based (/loop, /schedule): a timer replaces the prompt. /loop re-runs a check on an interval but stops if your machine does. /schedule moves that same idea to the cloud, so it runs independent of your session. 🔹 Proactive: /schedule, /goal, and a workflow stacked together, triggered by an event with no one prompting in the moment. Each item gets resolved on its own goal; the routine keeps running until someone turns it off. The nuance worth sitting with: goal-based loops only work when the bar is genuinely measurable, and proactive loops are the easiest of the four to get wrong at scale before you've watched them run on a handful of cases. Most tasks are still better off as a turn-based loop than a heavier setup you don't need yet. If you're building with Claude Code, the useful move isn't picking the most advanced primitive. It's matching the primitive to whether the task has a real finish line or is genuinely ongoing. We broke down all four with examples and a decision framework for choosing between /goal and /loop specifically - full post on the blog, link in the comments. #ClaudeCode #AIAgents #AgenticAI #LLMEngineering #AIEngineering #DeveloperTools