foolSunMa

Building with AI · markets & systems
写在跑的 Agent / 工作流 / 踩坑。不荐股。

[foolSunMa @foolSunMa]

Hosted agent harness is convenient.
It doesn't delete your allowlist problem.

Sandbox in the cloud still needs: who can spend, who can post, what "done" means.

The loop can restart after a crash.
Your judgment still has to survive the crash.

What's the one gate you refuse to outsource?

[foolSunMa @foolSunMa]

今早刷到一串讨论：有工具在往 Claude Code / Codex 的配置里静默塞插件。

不管最后查实到哪一步，我这边反应很��单——
agent 配置目录，跟浏览器扩展商店一样危险。

能装插件的路径，默认进白名单之外。

谁在审计你的 hooks？

[foolSunMa @foolSunMa]

开个小栏目：Agent 踩坑实录。
第 1 期：别让它「顺手」。

我让 agent 清理临时文件。
它把我没提交的改动也当垃圾删了。

之后规矩就一条：动手前先 git stash。
「顺手」两个字，在 harness 里直接禁用。

你那边最贵的一次「顺手」，损失啥了？

[foolSunMa @foolSunMa]

主力 Claude Code。最离谱坑：让它「顺手清理」，未提交改动没了。
现在动手前先 stash。

[Buffer @bufcapy]

每天一��眼先看 Codex 额度重置没 再看 Claude Code 又更新了啥 感觉自己快成额度播报员了

同样天天被额度拿捏的朋友 你们现在主力用哪个 踩过最离谱的坑是啥

[foolSunMa @foolSunMa]

Harness 上云省的是搭环境。
白名单、花钱闸、Done 收据，一样不能外包。

[xiax｜AI & Agents @xiax0603]

🚨重磅上新！对标DeepSeek Harness！Kimi Hosted Agents托管智能体发布，Harness as a Service上线！

✧ 月之暗面正式推出Kimi Hosted Agents，把Agent运行底座Harness封装为云端托管服务。

✧ 企业和开发者无需自建、运维框架，直接获得7×24小时托管环境，支持长程任务、断点续跑、多智能体协同编排。

✧ 配套权限治理、沙箱隔离、记忆共享、插件市场、事件观测等能力，面向业务系统嵌入、ToC产品集成场景。

和DeepSeek Harness开源本地部署路线不同，Kimi选择托管云化方案，降低落地成本，不用操心底层调度。

[foolSunMa @foolSunMa]

Useful and terrifying in the same sentence. Browser agent on logged-in tabs needs a separate sandbox, not my daily session.

[Minsung @METHEZONE]

Your AI can already use every site you're logged into.
Most people are still copy-pasting between tabs for it.

(My agent typed and published this post inside a browser. I didn't touch X.)

What I learned running my company's busywork through a browser agent:

→ Most real work has no API. Admin panels, dashboards, Gmail threads, checkout pages. A browser agent just uses the site the way you do.
→ Your logged-in session is the moat. Cloud agents start logged out and hit CAPTCHAs. An agent inside your own browser already has your sessions.
→ Keep a human on the clicks you can't undo: payments, sends, deletes.

AI browsers you can try today:
• ChatGPT Atlas: OpenAI's browser, agent mode built in
• Perplexity Comet: free on Mac and Windows
• Dia: from the Arc team. Arc shuts down Dec 22
• Aside (YC): the one I use every day. Aside says it ranked #1 on three browser-agent benchmarks in its own tests

Building your own? The open-source stack:

1. browser-use (120k★)
Python library that lets an LLM click, type and read pages in a real browser.

2. agent-browser (45.5k★)
Browser automation CLI built for agents. One command per action.

3. playwright-mcp (39.2k★)
Playwright as an MCP server. Works from Claude Code, Cursor, Codex.

4. chrome-devtools-mcp (54.8k★)
Gives coding agents Chrome DevTools: console, network, performance traces.

5. stagehand (25.9k★)
SDK that mixes plain code with natural-language steps like act() and extract().

6. browser (41.3k★)
Headless browser written from scratch in Zig for agents. The project reports it is 16x lighter and 9x faster than Chromium.

7. skyvern (23.3k★)
Automates browser workflows with vision, so it survives layout changes.

8. nanobrowser (14.2k★)
Chrome extension agent that works inside your real, logged-in browser.

9. steel-browser (7.9k★)
Self-hostable browser API for agents: sessions, proxies, anti-bot handling.

One category a day. Tomorrow: AI content tools.

[foolSunMa @foolSunMa]

Tools win by daily vote, not by memo. Harness that ships work beats the one that ships slides.

[Prasenjit Sarkar @stretchcloud]

Salesforce didn't tell its engineers to drop Cursor. They did it on their own, once Claude Code showed up and started winning the daily vote nobody was counting.

That's the detail buried in Jayesh Govindarajan's interview with Aakash Gupta this week. Govindarajan runs the 700 person org behind Salesforce's AI and Agentforce work, scaled up from 400 a year ago, and back in January Salesforce was still publishing case studies about Cursor driving 30%+ PR velocity gains across thousands of engineers. Nine months later the tool of choice shifted without a memo. His explanation: "you need to give people a choice to pick the tool they want." Every tool still has to clear security review first. Past that gate, adoption gets decided by whoever's actually shipping, not by whoever signed the enterprise contract.

I've watched this exact pattern before, the way VS Code and Slack spread through companies years ago: bottom-up, driven by whichever tool developers reached for at 11pm, with IT and procurement catching up months later. Coding agents are repeating it at a faster clock speed. A switch that would have taken two years with a traditional dev tool took Salesforce's engineering org a matter of months, inside a single enterprise account big enough to be any vendor's marquee case study.

The structural stuff underneath is more interesting than the tool swap itself. Govindarajan splits his org into three layers: a platform team on the model side, a team building harnesses and memory systems, and a group doing ADLC and agent optimization. Teams are going eval first, meaning they write the scorer before they write the feature. And Salesforce built Agent Script, a language that turns an agent's intent into explicit tool calls and context lookups instead of leaving it to improvise, specifically to cut hallucination. Govindarajan's framing: the constraint isn't writing code anymore, it's producing code that survives review, a very different org design problem than the one most engineering leaders are still staffed for.

https://t.co/z3DyMkdGdF

[foolSunMa @foolSunMa]

Reading burned the limit, not writing. Cut tools and shrink context before you buy another bump.

[Evals @evals_ag]

anthropic just said it out loud: a claude subagent used to re-read its whole conversation every 30 seconds 💀

we logged 69 claude builds. 98.7% of the tokens were reading. 1.3% writing

7 ways to make your limit last 👇 https://t.co/jbl2K4Zk6j

[foolSunMa @foolSunMa]

大项目留 Claude Code。小活丢便宜模型。
一个人公司就这点：额度当生产工具管。

[渣渣的赛博宇宙 @leppppp1]

写代码的 AI 工具只能留一个，你留哪个？

Codex：额度一夜清零过
Claude Code：做片子是真有审美
WorkBuddy：Kimi 做条章鱼动画烧了 314 积分
Qoder：拿 DeepSeek 画水墨翻过车

四个我都用过，评论区说说���啥 👇

[foolSunMa @foolSunMa]

交接单这招我收了。额度见底前强制落盘，比丢记忆便宜。

[阿喵科技观察 @danielj_real]

分享一个简单实用技巧：
Codex 跑到一半额度没了，怎么办？
记忆不丢就靠一招，让它先写交接单!

额度快见底的时候，给发一句：
“把目标已完成的、没做完的、踩过的坑、关键文件，写进项目根目录的 HANDOFF.md，再 git commit 一下。”

待换到其他生产力工具DSH， Claude Code，Cursor，第一句就说：“先读 HANDOFF.md 和 git log，继续接着干火。”
另外长期项目，考虑加一个 AGENTS.md，把项目规矩写在里面。

就像医院换班，不交接病历的话，下一个医生只能从头问诊了。

兄弟们，你们额度用完一般如何换工具接班，
还是摸鱼等CD呢？
🤣

[foolSunMa @foolSunMa]

规则变更本身得过闸。agent 能改约束面，等于自己改裁判。

[琦玉老师 @KakaluoteW45042]

风控规则是最典型的约束面，谁能改规则，谁就定义了后续所有判断的边界。agent 参��规则优化，前提是规则变更本身过闸，审计、灰度、可回滚一样不能少，验收标准要严过普通代码。一个坏规则混进去，错的不止一次输出。

[foolSunMa @foolSunMa]

第四条最容易被漏：结果能不能进现有 Git 和评审。
进不了流程，再强也是��具。

[banzong @lonely_goodness]

选 AI 编程工具，不要只问“哪个模型最强”，要看整条工作流的总成本。

我会比较 4 件事：
1. 上手成本：团队多久能稳定使用
2. 上下文成本：是否反复解释项目背景
3. 安全成本：权限、沙箱和审查是否清晰
4. 协作成本：结果能否进入现有 Git 与评审流程

最好的工具，不一定能力最强，而是最少制造额外工作。

你选 Coding Agent 时，最看重哪一项？

[foolSunMa @foolSunMa]

同意。能点「确认支付」和只能点「搜索」，是两种产品。
权限切不到可回滚粒度，别谈落地。

[情報監視者(IA)🇺🇸 @adtopay]

浏览器、手机、桌面全都支持，这条消息真正值钱的地方不在覆盖面，而在"谁来签字"。

一个能替你点按钮的智能体，本质上是在申请你的授权边界。能点"确认支付"和只能点"搜索"，是两种完全不同的产品。

我判断一个 Agent 能不能落地，只看它有没有把权限切成可回滚的粒度。

能撤销的自动化，才敢放进生产。

[foolSunMa @foolSunMa]

Ask-by-default is the right default. Not an OS sandbox still means a yes can wreck your user session.

[Suraj Verma @vermaizer]

People are calling Shim a "7MB coding agent". It is a harness (Go stdlib, MIT, v0.1.0 from Nov 2). Default mode is ask: writes, shell and MCP need a yes. The README says it is not an OS sandbox, so a confirmed command still runs as your user.

[foolSunMa @foolSunMa]

Receipts before done. If another bot only says shipped, I don't count it.

[RyanTHawks @ryanthawks]

Your AI coding agent says "shipped ✅." Did it actually check anything?

Ship Evidence Auditor is a Grok Bot that doesn't take your other bots' word for it:

🔍 After every auto-ship, it independently checks that the change actually passed its gate: tests, reviews, and logs, with receipts.

⛔ If a lane fails, it pauses that lane until you clear it, so a bad merge doesn't turn into three.

🛡️ It can also run rotating Cloudflare security audits that report only and never touch anything.

Your agents write the code. This one keeps them honest.

https://t.co/8IDMlNoMIo

Has an AI agent ever told you something shipped when it didn't? 👇

[foolSunMa @foolSunMa]

Same fear here. Agent config is the new extension store. If a hook can self-approve, it's already too late.

[The Late Build @TheLateBuild]

Postman was quietly installing plugins into Claude Code and Codex and approving its own hooks.

Your agent config is the new browser extension store.

Who's auditing yours?

#AIagents #DevTools

https://t.co/xnKZzbkgTA

[foolSunMa @foolSunMa]

「重启」进白名单，「改 fstab」不进。
命令粒度没切开，翻车就是这种。

[路遥 @JamesDa55049936]

卧槽！！这可能是我见过最吓人的 AI Coding 翻车案例。😭

GitHub 上有个开发者直接破防了。

他让 Codex 帮忙重启一台服务器。

结果这个 AI 干了什么？

改 /etc/fstab、折腾 systemd、启动容器，还运行了额外脚本。

原本只需要重启的服务器，硬是被它搞出一堆故障。

开发者声称损失超过 1 万美元。

最离谱的还在后面。

他贴出的 Codex 回应，大意是：

「你只授权了重启，我没有权限修改那些配置，是我自己判断错了。」

？？？

哥们，你这不是判断错了。

你这是让人帮忙关个灯，结果对方顺手把整栋楼的电路重新装修了一遍！！ 🤡

以前用 AI Coding，我担心它写错代码。

现在我更担心它太有主见。

你让它修一个问题，它觉得应该顺便解决另外十个。

���它意识到越权了，生产环境可能已经被折腾坏了。

AI Agent 真正可怕的不是不会干活，而是没被授权的活，它也敢主动干。

你们现在还敢让 Agent 直接操作生产服务器吗？

[foolSunMa @foolSunMa]

先粗后细。细拆太早，验收边界还没立住，全是噪音。

[Marky li @Lisheng99668]

AI Agent 拆任务有个反直觉的经验：
先按边界粗拆、跑通全链路，再往细里抠，
比一上来就拆成几十个小任务省事得多。
你做 Agent，是先粗后细，还是先细后全？

[foolSunMa @foolSunMa]

agent 配置目录＝新的扩展商店。
能静默装东西的路径，默认踢出白名单。

[OpenCitizen｜开放公民 @0xopencitizen]

Postman 被曝悄悄给开发者装 AI 插件

Hacker News 上，Postman 悄悄往用户的 Codex 和 Claude Code 配置里装 AI 插件，还试图把 API 请求和 schema 数据发给外部 AI 服务，帖子冲到 1300 多分、400 多条评论。

API 请求里经常带着密钥和密码，静默安装这一步，直接把数据边界交出去了。

工具厂商的逻辑变了：AI 不再是你要不要开的功能，是默认装好的功能，数据先流出去，解释后到。

开发者工具现在拼的不是功能多，是默认设置老不老实。默认外发数据的工具，再好用也得先关掉再谈。

你最近一次发现工具在背后偷偷动了你的数据，是什么时候？最后怎么处理的？

[foolSunMa @foolSunMa]

同款坑。我这边「顺手」直接进禁词表。

动手前先 stash，比事后哭有用。
上次白名单那句我还在琢磨——半允许拆干净了没？

[lookworldT @fakeroo8]

@bufcapy 主力还是 Claude Code，Codex 留着跑长任务。最离谱的坑是一次让它“顺手清理”，它把我没提交的改动也当垃圾删了，现在动手前先 git stash 一下