WorkBuddy 都帮你们做了哪些工作？我把他放到了竞技场上厮杀...

{"type":"doc","content":[{"type":"paragraph","attrs":{"id":"6e8a22ee-d020-4072-9693-ed2f34ec9ece","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"最近我在尝试一个比较有意思的玩法：不是让 AI 直接聊天，也不是让 AI 生成一段静态代码，而是让 AI 编写一个可以真正参与对战的 Agent 策略代码，然后把它提交到一个竞技场里，看它在回放中如何移动、判断、攻击和求生。"}]},{"type":"paragraph","attrs":{"id":"d0b69376-1b7d-4b70-8253-1d24c6e9e8b5","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"模型选择的是 GLM 5.2，让它帮我编写 AgentDuel.app 上的 Agent 对战代码，。实际体验下来，最直观的感受是：WorkBuddy 很适合这种“读规则、理解接口、生成代码、再根据结果迭代”的开发场景，而 GLM 5.2 在策略理解和代码生成上的表现也比我预期更强。"},{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"AgentDuel.app 的玩法可以简单理解为“代码之间的竞技场”。玩家不需要手动操作角色，而是提交一段 Agent 策略代码，让这个 Agent 根据当前战场状态自主决策。比如判断敌人距离、选择移动路线、释放技能、进入草丛躲避、追击残血目标等等。最后，网站会根据战斗日志生成完整回放，你可以像看一场小型 AI 机器人比赛一样观察它们的决策过程。"}]},{"type":"image","attrs":{"id":"4a795fa2-c8a1-4282-b168-9a8536abb05e","src":"https://developer.qcloudimg.com/http-save/audit-5334288/705550a088f005753ebee67bfebff031.png","extension":"png","align":"center","alt":"","showAlt":false,"href":"","boxShadow":"","width":583,"aspectRatio":"1.362150","status":"success","showText":true,"isPercentage":false,"percentage":0,"isHoverDragHandle":false}},{"type":"paragraph","attrs":{"id":"6d82fb41-d8e1-4d4d-b0b0-f5f5595a673b","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"我这次的流程很简单：先把 AgentDuel 的规则和接口文档交给 WorkBuddy，然后让它基于死斗模式生成一份可提交的 Agent 代码。相比自己从零开始写，WorkBuddy 的优势在于它能比较快地把文档里的约束转成可运行的代码结构，比如如何读取战场状态、如何判断技能冷却、如何选择目标、如何在不同距离下切换攻击或移动策略。"}]},{"type":"image","attrs":{"id":"84026631-c1c1-47ee-a1b8-9090e071818e","src":"https://developer.qcloudimg.com/http-save/audit-5334288/d7add073e1dd002cfdfdcd6dd68e44b0.png","extension":"png","align":"center","alt":"","showAlt":false,"href":"","boxShadow":"","width":785,"aspectRatio":"1.736726","status":"success","showText":true,"isPercentage":false,"percentage":0,"isHoverDragHandle":false}},{"type":"paragraph","attrs":{"id":"3e09b7b6-91f5-49be-b58f-17d1f8b232fb","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"比较让我惊讶的是，GLM 5.2 生成的策略并不是简单地“看到敌人就冲上去打”。它会尝试根据距离、血量、技能状态来做一些基础判断。虽然还谈不上非常复杂的博弈，但已经能体现出一定的战术倾向：能打就打，打不到就靠近，危险时尝试拉开距离或寻找更好的位置。这种策略在回放里看起来很有意思，因为你能清楚看到代码决策如何影响战斗结果。"}]},{"type":"image","attrs":{"id":"1ee2e063-7184-4a30-872d-9c49e032956b","src":"https://developer.qcloudimg.com/http-save/audit-5334288/c23f7fceaec5ff0f0422069d15cc46ea.png","extension":"png","align":"center","alt":"","showAlt":false,"href":"","boxShadow":"","width":1100,"aspectRatio":"1.709459","status":"success","showText":true,"isPercentage":false,"percentage":0,"isHoverDragHandle":false}},{"type":"paragraph","attrs":{"id":"9e34a6a5-041a-437d-b129-b86c320f60d0","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false}},{"type":"paragraph","attrs":{"id":"be2b4aad-e79b-4982-af29-c97cee7cbd00","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"把代码提交到 AgentDuel.app 之后，就可以直接发起对战。网站会运行双方 Agent 的策略代码，并生成一场完整的战斗回放。这个过程最有趣的地方在于，你不是在看一段预设动画，而是在看两个由代码驱动的 Agent 互相博弈。每一次移动、攻击、技能释放，背后都对应着 Agent 在当前局面下做出的判断，查看我的这场对战回放。"}]},{"type":"image","attrs":{"id":"f8155324-e289-4d19-ad24-3e1d3c85a425","src":"https://developer.qcloudimg.com/http-save/audit-5334288/f5c2759e2599fdb9cc06e3fd1857bc4e.png","extension":"png","align":"center","alt":"最后我的Agent击杀了对方的m3 Agent","showAlt":true,"href":"","boxShadow":"","width":1100,"aspectRatio":"1.118349","status":"success","showText":true,"isPercentage":false,"percentage":0,"isHoverDragHandle":false}},{"type":"paragraph","attrs":{"id":"a5887f2c-6556-4b3a-acc1-9f418dea0a7a","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"从开发体验上看，WorkBuddy 在这类场景里非常省事。过去如果要实现一个能对战的 Agent，我可能需要先读完规则，再手写接口调用，再一点点调试逻辑。但现在更像是把 AI 当成一个策略代码搭档：我负责描述目标、提供规则、观察回放和提出改进方向，WorkBuddy 负责快速生成和修改代码。这个协作方式明显提高了试错速度。"}]},{"type":"paragraph","attrs":{"id":"2c0d14b2-f84c-4413-94dd-9971422df983","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"当然，AI 生成的 Agent 代码不一定一次就完美。真正好玩的地方也在这里：你可以看完回放后继续让 WorkBuddy 优化策略，比如“战士不要无脑追击，要优先卡距离”“残血时尝试往草丛移动”“技能冷却时不要浪费行动力”等等。每一轮迭代都会让 Agent 的行为更接近你想要的战术风格。"}]},{"type":"paragraph","attrs":{"id":"53b260b2-3979-4495-8465-9eacbd2f9c66","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"这也是我做 AgentDuel.app 时最想验证的一点：当 AI Coding Agent 变得越来越强之后，游戏里的“玩家操作”也许不一定只发生在键盘和鼠标上，也可以发生在提示词、策略代码和模型能力之间。你不只是控制一个角色，而是在训练、调整、比较不同 Agent 的决策方式。"}]},{"type":"paragraph","attrs":{"id":"79d6b1da-fb3b-416f-a945-314ecbb51d3b","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"这次用 WorkBuddy GLM 5.2 写 Agent 对战代码的体验，整体比我预期顺畅很多。WorkBuddy 负责把想法快速落到代码里，AgentDuel.app 负责把代码放进竞技场里验证结果，两者结合起来，有一种“让 AI 写代码，再让代码上场打架”的感觉。"}]},{"type":"paragraph","attrs":{"id":"4d6d0f69-3ed4-486d-a6b8-97a8439e42ab","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"如果你也对 AI Agent、代码对战、策略模拟或者 AI 编程工具感兴趣，可以试试这种玩法：让 WorkBuddy 写一个 Agent，把它提交到 AgentDuel.app，然后看看你的 AI 代码在竞技场里到底能打成什么样。"}]}]}","createTime":1783320933,"ext":{"closeTextLink":0,"comment_ban":0,"description":"","focusRead":0},"favNum":0,"html":"","isOriginal":0,"likeNum":0,