[{"data":1,"prerenderedAt":46},["Reactive",2],{"news-story-news-20261025-x-discussion-thorsten-ball-how-i-work-with-agents":3},{"id":4,"title_zh":5,"title_en":6,"summary_zh":7,"summary_en":8,"takeaways_zh":9,"takeaways_en":15,"content_zh":21,"content_en":22,"external_links":23,"source":24,"category":31,"impact_score":32,"views_count":33,"likes_count":34,"related_entities":35,"tags":40,"image_url":45},"news-20261025-x-discussion-thorsten-ball-how-i-work-with-agents","X 热议：Thorsten Ball 发长文《How I work with Agents》，分享 15 条 AI 编程 agent 工作法，近 4 千收藏","X discussion: Thorsten Ball's \"How I work with Agents\" lists 15 habits for coding with AI agents and draws nearly 4K bookmarks","开发者 Thorsten Ball（@thorstenball）在 UTC 2026-10-25 15:58 发布 X 长文《How I work with Agents》（[原帖](https://x.com/thorstenball/status/2114386050638336082)），总结他「截至 2026 年 10 月」和编程 agent 协作的做法：只用最新最强的模型、提需求只说做什么不说怎么做、让 agent 看到从部署到线上日志的整个研发流程、要求 agent 拿出「无可辩驳的证据」、只审查影响面大的代码等。抓取时约 1,935 赞、143 转发、92 回复、3,981 收藏、13.7 万浏览，并进入 X 热门话题。这是个人经验分享，观点按观点呈现，不是经过验证的结论。","At 15:58 UTC on 2026-10-25, developer Thorsten Ball (@thorstenball) published an X article, \"How I work with Agents\" ([post](https://x.com/thorstenball/status/2114386050638336082)), summarizing how he works with coding agents \"as of October 2026\": use only the latest and strongest models, prompt for what you want rather than how to do it, make the whole development lifecycle from deploys to production logs visible to the agent, ask agents for \"irrefutable proof\", and review only code with a high blast radius. When we pulled it, the post had about 1,935 likes, 143 reposts, 92 replies, 3,981 bookmarks and 137K views, and it reached X's trending stories. This is a personal account of practice presented as opinion, not a verified conclusion.",[10,11,12,13,14],"主帖：[原帖](https://x.com/thorstenball/status/2114386050638336082)，UTC 10-25 15:58 发布，抓取时约 1,935 赞、143 转发、92 回复、3,981 收藏、13.7 万浏览；收藏数约为点赞的 2 倍，说明很多人把它当清单存下来","模型选择：作者主张只用最新最强的模型，理由是模型错 10% 时你不知道是哪 10%，只能反复重试","提示方式：只描述想要的结果，例如「把字节码编译器和 VM 做到尽可能快，不到 2 倍提速不要停」","验证与审查：让 agent 提供端到端演示、视频、前后对比数字作为证据；作者自己只审查不到 5 个核心文件和数据库迁移等影响面大的改动","工作节奏：别等 agent，让 agent 等你；作者称定时、唤醒、心跳类自动化是「杀手级功能」，并认为上下文压缩已足够好，不必手动管理 token",[16,17,18,19,20],"Main post: [post](https://x.com/thorstenball/status/2114386050638336082), published 15:58 UTC on 10-25, with about 1,935 likes, 143 reposts, 92 replies, 3,981 bookmarks and 137K views when we pulled it; bookmarks are about twice the likes, suggesting many readers saved it as a checklist","Models: the author says to use only the latest and strongest models, because when a model is wrong 10% of the time you do not know which 10%, so you end up retrying","Prompting: describe the outcome, for example \"Make the bytecode compiler and VM AS FAST AS YOU CAN. Do not stop until it's 2x faster.\"","Verification and review: ask agents for end-to-end walkthroughs, demo videos and before/after numbers as proof; the author personally reviews only high-blast-radius changes such as fewer than 5 central files and database migrations","Workflow: do not wait for agents, let them wait on you; the author calls schedules, wake-ups and heartbeats a \"killer feature\" and says context compaction works well enough that manual token management is wasted effort","### 核心背景与行业痛点\n随着编程 agent 能连续执行长任务，很多开发者卡在「怎么用」而不是「能不能用」：提示写得太细、反复来回微调、担心代码变成一堆难维护的「slop」。UTC 10-25 15:58，开发者 Thorsten Ball（@thorstenball）把平时私信给朋友的建议整理成 X 长文《How I work with Agents》（[原帖](https://x.com/thorstenball/status/2114386050638336082)），开头注明是「截至 2026 年 10 月」的做法。文章很快进入 X 热门话题，抓取时收藏数（3,981）约为点赞数（1,935）的两倍。\n\n### 架构亮点与底层机制\n文章的核心思路是「让 agent 看见现实」。作者认为，真正减少幻觉的是工具调用：agent 看不到的东西就会编造，所以要让它端到端运行自己写的东西，读线上日志、部署、观察发布、切换功能开关、处理 bug 报告。作者提到团队为 bug 报告、给用户发邮件、系统健康检查和授予功能开关做了自定义插件，并以 Amp 为例说明会话记录（thread）与 git 提交关联后可以成为团队记忆。他还建议用对照实验评估 AGENTS.md、skills 等知识文件是否真的被 agent 使用：开 5 个互不知情的 agent，分对照组和实验组比较。\n\n### 权威 Benchmark 与实测跑分对比\n本条是 X 热议和个人经验分享，没有公开跑分，文中也没有给出可复现的实验数据。可核对的只有帖子互动数据：约 1,935 赞、143 转发、92 回复、3,981 收藏、13.7 万浏览。文中「编译器提速 2 倍」只是提示示例，「有的会话 46 条消息、有的 3,714 条」是作者自述，均非基准测试结果。\n\n### 开发者实战落地与开箱指南\n可以直接试的几条：一，把「怎么做」换成「要什么结果」，并给出可衡量的目标；二，不要和 agent 一轮轮试错，而是让它一次生成 20 个变体并做个小页面方便比较，排查问题时让它先提出 4 个假设再逐一验证；三，在 AGENTS.md 里写明什么不算证据，要求 agent 用端到端演示或前后数字证明完成；四，把任务做大、改成异步批量，并用定时检查代替盯屏。作者也承认本地开发环境在多个 agent 同时跑端到端测试时容易资源冲突，远程机器更合适。是否全盘照搬，建议结合团队的代码审查要求判断，尤其是「不再在意代码重复和变量命名」这类观点争议较大。","### Background and the problem\nAs coding agents become able to run long tasks on their own, many developers are stuck on how to use them rather than whether they can: over-detailed prompts, endless back-and-forth tweaking, and worry that the codebase will fill up with unmaintainable \"slop\". At 15:58 UTC on 10-25, developer Thorsten Ball (@thorstenball) turned the advice he usually sends friends in DMs into an X article, \"How I work with Agents\" ([post](https://x.com/thorstenball/status/2114386050638336082)), labeled as his practice \"as of October 2026\". It quickly reached X's trending stories, and when we pulled it the bookmark count (3,981) was about twice the like count (1,935).\n\n### How it works\nThe core idea is to let the agent see reality. The author argues that what really stopped hallucinations was tool calls: if the agent cannot see something, it makes it up, so it should run what it builds end to end, read production logs, deploy, watch rollouts, flip feature flags and triage bug reports. He says his team built custom plugins for bug reports, emailing users, system health checks and granting feature flags, and uses Amp as an example of how threads linked to git commits become team memory. He also suggests evaluating whether knowledge files such as AGENTS.md and skills are actually used, by spawning five blind agents split into control and treatment groups.\n\n### Benchmarks and measured results\nThis is an X discussion and a personal account of practice; it has no public benchmark and no reproducible experimental data. The only numbers we can check are engagement: about 1,935 likes, 143 reposts, 92 replies, 3,981 bookmarks and 137K views. The \"2x faster compiler\" line is a prompt example, and \"some threads are 46 messages, some are 3,714\" is the author's own description; neither is a benchmark result.\n\n### What developers can do now\nA few things you can try directly. First, replace \"how to do it\" with \"what result I want\" and give a measurable target. Second, instead of trial and error turn by turn, ask the agent to generate 20 variations plus a small page to compare them, and when debugging ask for 4 theories and have it check all of them. Third, write in AGENTS.md what does not count as evidence, and require end-to-end demos or before/after numbers as proof of completion. Fourth, give agents bigger tasks, work asynchronously in batches, and use scheduled checks instead of watching the screen. The author also admits local development gets resource conflicts when several agents run end-to-end tests at once, so remote machines fit better. Whether to adopt it wholesale depends on your team's review standards; views such as no longer caring about duplication or variable names are the most contested.",[],{"platform":25,"author_name":26,"author_handle":27,"author_avatar":28,"tweet_url":29,"published_at":30},"x","Thorsten Ball","@thorstenball","https://pbs.twimg.com/profile_images/2073288299528024064/3itOM_L5_normal.jpg","https://x.com/thorstenball/status/2114386050638336082","2026-10-25T15:58:12.000Z","tooling",82,91,0,{"models":36,"agents":37,"vendors":39},[],[38],"Amp",[],[41,42,43,44,26],"X热议","AI编程","编程Agent","最佳实践","https://pbs.twimg.com/media/HVfOvcmXkAAhUhz.jpg",1793718871593]