Leo

AI Agents
Check my personal site, i promise i will update 😉

[Leo @YangLi_leo]

I definitely believe cloud agents are the next problem that needs to be solved, since they will also define what the next generation of cloud infrastructure looks like. I’ve spent the last few months getting started on this with Tetral, which is fully open-source: https://t.co/EJZuHUs9XP

I think that in this era, open-sourcing the code isn’t enough anymore. Or I should say, letting people know how every decision is made has become more important than ever. So I strongly recommend reading this blog post: https://t.co/7FjjRv8FSv

I’m grateful to the authors of every blog post I mentioned in the piece. This wouldn’t have been possible without your work.

I want to keep exploring the next generation of cloud agents. If you’re interested in this project, please don’t be shy about joining me on GitHub!

[Dmitriy Kovalenko @neogoose_btw]

I also have a little announcement for DevDay. I asked Tibo to include it in the presentation but he didn't read my slack DM 😭

My first onboarding project is merged. I fixed this issue once and forever! I reduced amount of disk writes by 50% and the size of stored files by 4x

[Leo @YangLi_leo]

@Yuchenj_UW so that's why oai's dev day is trying so desperately to cut developers' usage and burn some more money to normal users, at the end of the day, they wanna to be a giant more than coding agent

[Leo @YangLi_leo]

@nichochar i just wonder what's your take on space, which is a potential competitor with Notion

[Leo @YangLi_leo]

@fiapp_pro 6.1 sol和astra相比有啥优缺点？

[Leo @YangLi_leo]

@flrande 这要追的话其实最早的是A\的managed agent sdk，这也是我的项目选择用它作为interface的原因，设计的时候考虑得很周到，版本控制，vault等等

[Leo @YangLi_leo]

所以确实只要做的足够慢，就不用加花里胡哨的东西了🤣

当时做自己的harness的时候就已经发现模型虽然有时用todo list，除了ui上一目了然之外，对最后结果已经没啥影响了，当机立断选择没做lol

[yetone @yetone]

是的，现在还有 todo list 的 harness 基本上都是上一代产品

[Leo @YangLi_leo]

@arkuy99 acp希望定义行业标准，可能看到mcp的成功感觉自己行了，没想到各个原厂看都不看，最后变成了追着别人修自己的接口，这个loop就循环下去了

[Leo @YangLi_leo]

Why? I think they don’t want to be remembered just as a chatbot forever. They want to be deeply integrated into users’ daily lives, which means both the product form and their marketing strategies have to evolve. As a developer, I hate seeing my Codex credits burn faster than ever, but honestly, it’s the right move.

[Tibo @thsottiaux]

Launching a new Codex Cloud, much improved from last year. With configurable cloud environments, it’s impossible to go back to building on your laptop once you’ve taken the time to configure it.

Agents API, the same tech powering all our cloud agents, including dots, is also now in preview and supports computer use. It allows you to build the same incredible products we are making available today.

[Leo @YangLi_leo]

没啥说的了，精准预测，OpenAI dots也是完全基于自家的Agent API构建的，cloud agent api一定会替代传统的各类精细控制的api打包出售

从Managed Agents SDK出来的第一天，它设计的全面性，几乎覆盖了所有痛点，几乎对我来说是一个完美的解，A\今年最大的前瞻是在三月就能做出开这么完备的体系，同时这也是claude code之后的另一个爆点

当然tetral会持续适配和优化Managed Agents SDK compatible端口，在目前包括codex在内的开源的agent也开始把很多工具迁移到云端，agent之间加密通讯，保持一个开源的路径给大家探索还是很重要的

[Leo @YangLi_leo]

I think when Anthropic released Managed Agents, most people dismissed it as no big deal. But trust me: as all kinds of local workflows migrate to the cloud, scaling agent teams in a stable environment is definitely going to be a huge deal

[Leo @YangLi_leo]

看来OpenAI是把最近市面上火的东西全部做了一遍，jev也没能幸免

[Tibo @thsottiaux]

Decisions API, for lightning fast constrained decision making powered by Luna. Supports visual inputs, and tuned to be able to make decisions in less than a few hundreds of milliseconds end to end.

[Leo @YangLi_leo]

@Zen_with_AI omg，这和我的想法可以说是一模一样了，已经在往这个方向做了，有兴趣的话可以看看： https://t.co/Vkr6KdZhkM，我是几乎完全基于这个理念写了一套自己的cloud agent runtime

[dax @thdxr]

for arguments sake let's accept that spending a lot of tokens is truly more productive

we're in a weird stage where early stage startups are disadvantaged. they cannot afford to do this

it's reverting back to the days when you had to raise money to buy servers to even launch

[Leo @YangLi_leo]

@truerickc137 我是合成数据门外汉，只是当时有过一点想法，但真的就停留在想法阶段lol因为从来参与过training infra的搭建
看到DSec之后还挺意外的，他们把agent loop放进了sandbox，就是为了保证数据在sandbox里，我的理解是

[Leo @YangLi_leo]

@KakaluoteW45042 兜兜转转agent只要还是一个软件，那就不可能超越之前的设计哲学，说多了都是泪😂

[Leo @YangLi_leo]

@An_yhl 这就是正在打磨的升级版，需要大量的测试，两周内会出一个非常详细的blog和源码推送，之后也会有大量的feature偏向用户侧的而不仅仅是架构侧的

[Leo @YangLi_leo]

这还是得看看我的Tetral：https://t.co/LK1oyrpphw，在开源版本里做的最好的，目前找不到更完整的了，当然最近在筹备一场大更新，迈向多节点

���全自研的durable agent runtime，因为目前没有agent适合直接部署在云端，都无法匹配agent out of sandbox

[Yifan Xu @yifanxu_ephai]

今天试用 Cue 时，我又想起最初做 sandbox 时的一个问题：

为什么我们总默认“一个 Agent 对应一台 computer”？无法实现以下两种场景。

场景 1: 我希望 Agent A 同时使用两台电脑：在电脑 1 上改代码、跑测试，在电脑 2 上查资料、验证产品流程。

场景 2: 两位 Agent 可以协作使用同一台电脑。Agent A 写代码，Agent B 负责review。

tbh 我不太看好这种runtime one-on-one绑定的模式。

未来一定agent比人多，比电脑多，比可以启动的vm，服务器多得多。各大厂商似乎还没有意识到这一点。

[Leo @YangLi_leo]

感谢推荐！之前不是很主动在各种社媒上说话因为感觉看看就行了，但这次cloud agent集中爆发让分享欲压不住了lol

agent发展到现在已经完全不是一个简单的问题了，是由无数底层优化堆起来的，现有的infra远远做不到能充分承载agent不断发展的能力，同时agent不��仅作为一个产品，本质上也要不断反哺模型的训练，像layerfs这些项目都是在这个路上重要的基石，这���是我这篇文章叫做the scaling problem的原因

目前tetral项目在筹备大升级，所以这两个星期没有提交新的代码，每一处改动都在斟酌，敬请期待！

[Yifan Xu @yifanxu_ephai]

非常推荐大家去看 @YangLi_leo 写的这篇文章: The Next Scaling Problem。前段时间也受到了群里大佬的高度赞赏。

文章里几个比较重要的章节和理念值得大家一看：

1. A computer should be something the agent calls, not somewhere the agent lives
2. Write ahead of execution
3. A durable participant
4. Scaling the system around intelligence

介绍了为什么要让 agent live outside of sandbox，把cloud agents存算隔离的架构解释得非常透彻，以及提供了架构如何enable multi-agent workspace 的非常好的insight。

https://t.co/u0fVnRBNYZ

[Leo @YangLi_leo]

@KakaluoteW45042 建议阅读一下这个blog，里面有很系统性的说明，当然还有些问题我目前也在测试，但是核心思路都已经分享出来了：https://t.co/Vkr6KdZhkM

[Leo @YangLi_leo]

@fiapp_pro 论openclaw的打包再出售