Arthur

[Arthur @Arthur_UFCAT]

@xueyu1125 主要是Jev虽然概念很好，但是模型推理能力真的不太行。。

[Arthur @Arthur_UFCAT]

@xueyu1125 As always... 用了开源项目而不感谢

[Arthur @Arthur_UFCAT]

@imwsl90 我还在用minimax 最近出的2.8 音频生成，还可以。 不知道有没有别的替代品？

我感觉minimax的方向就已经转到这块了

[Arthur @Arthur_UFCAT]

@realchendahuang 这个时候就要用到Personal Agent了，我现在会完全让Dot来帮我筛选有价值的推文 / 文章，如果质量过不了Agent的话我就直接忽略了🤣🤣🤣🤣🤣

[Arthur @Arthur_UFCAT]

@guansi 太有同感了，我最近越来越觉得，我们是最后一代程序员了，现在的新人就拿着AI生成的内容提交上来。你花大量时间跟他讲哪里哪里不对，他只会把你的意见原封不动的扔给AI。。。 所以，我花时间review帮助新人成长有什么意义吗。。。 还不如回头去弄pipeline期望有一天做到完全ai-native彻底摆脱带新人。

[Arthur @Arthur_UFCAT]

@FUCORY I'd have the harness record which task created each tab or worktree, then check for unfinished changes and other tasks still using it. That would give the user a concrete cleanup list to approve.

[Arthur @Arthur_UFCAT]

@huangtongxueh 验收报告里的每一项，可以直接链接到对应的代码和那次测试的证据。Review 的时候就不用再把需求、diff 和日志重新对一遍。

[Arthur @Arthur_UFCAT]

@JoshARosen Does the skill need to contain the decision logic? For something like lead prioritization, it could call a decision API with the current account context, while the proprietary criteria stay server-side.

[Arthur @Arthur_UFCAT]

@kunchenguid Have you tried using an empty contexted subagent to write test, we know an agent tends to trust its own decision in the long run though

[Arthur @Arthur_UFCAT]

@YinsenW_ Hermes的问题，代码结构设计得不好

[Arthur @Arthur_UFCAT]

@dair_ai Would be useful to compare full tool schemas vs on-demand loading in the same harness. Does the smaller context still save money once you count discovery calls and extra steps?

[Arthur @Arthur_UFCAT]

@eyad_khrais Approval should be enforced by the tool. A skill can change how the agent handles a case, but editing its instructions shouldn’t let it bypass approval.

[Arthur @Arthur_UFCAT]

@xiaomovps 我感觉把pi定位在后端服务里面可能更好。���。 毕竟toC的部分大多数用户还是希望有更傻瓜化的功能。。

[Arthur @Arthur_UFCAT]

@dabit3 I’d test a skill against specific tasks before keeping or deleting it. Reading a PDF, quoting a passage, and reusing a figure need different instructions. A skill can help with one and add noise to another.

[Arthur @Arthur_UFCAT]

@0xDevShah I'd freeze the model and test context/routing changes on held-out workflows first. If the same reasoning failures survive that, there's a much clearer case for owning the model.

[Arthur @Arthur_UFCAT]

@yifanxu_ephai 还有一层外部状态：邮件已经发出、CRM 已经改了，即使 session 和 workspace 一起回滚，也未必能撤回。这类动作的执行记录会怎么和 checkpoint 对齐？恢复时先核对实际效果，再决定重放还是跳过，感觉也需要进这个协议

[Arthur @Arthur_UFCAT]

@joseemv88 How do you handle rules like “material scope changes need client sign-off”? The units conflict can be compiled. I'd keep the scope-change judgment as an explicit, tested decision, with the approval requirement enforced separately at execution.

[Arthur @Arthur_UFCAT]

@HamelHusain For the handoff case, I'd label the decision point: given the context available then, should the agent act, ask, or hand off? Scoring only the final answer can hide the wrong branch. Each failure class needs its own cases before the agent starts optimizing it.

[Arthur @Arthur_UFCAT]

@Irvin279106651 看到了，但是还没试，主要是cc上面搭建了太多harness了，一时半会迁移不走……

[Arthur @Arthur_UFCAT]

作为重度的Claude使用者，长期保持2 - 3个20x max subscription账号进行开发，这一批核心Claude账号被封之后，我花了大约一周的时间找了一下替代品，我的核心发现和感受：

[Arthur @Arthur_UFCAT]

@jarvis11x SQQHRY