用客观校验约束 Grok Bot 工作流的 6 个开源仓库
作者借 AI Agent 花 5000 亿 token 重建游戏却先写错的案例，主张以可验证检查为准，并列出 LiteLLM、RouteLLM、Crawl4AI 等 6 个可围绕 Grok Bot 工作流测试的仓库。
AI AGENTS SPENT 500B+ TOKENS REBUILDING A GAME. THEIR FIRST MONTH OF CODE LOOKED RIGHT. IT WAS WRONG. In Maurice Heumann’s project, progress only became real when the code had to pass an objective check. I’d start there with Grok Bot, then add these 6 repos: 1/ LiteLLM · 61.6K⭐ Track model calls made by scripts working alongside the bot. → http://github.com/BerriAI/litellm 2/ RouteLLM · 5.6K⭐ Send those scripts’ requests to the model that fits each task. → http://github.com/lm-sys/RouteLLM 3/ Crawl4AI · 85.9K⭐ Turn documentation pages into clean, searchable text. → http://github.com/unclecode/crawl4ai 4/ LLMLingua · 6.8K⭐ Shorten large inputs, then check that the important details survived. → http://github.com/microsoft/LLMLingua 5/ Prefect · 24.0K⭐ Track each step and rerun the ones that need another attempt. → http://github.com/PrefectHQ/prefect 6/ Phoenix · 11.9K⭐ Inspect the workflow and spot work that keeps repeating for no reason. → http://github.com/Arize-ai/phoenix These aren’t the repos used to rebuild that game. They’re the tools I’d test around my own Grok Bot workflow. My takeaway: prompts, commits and hours spent aren’t proof of progress. Show me the check that confirms the result actually works.