张立行

观察人类，也折腾机器。
跨境电商 / AI & Agent / Enjoy building
生命的第一目标，是有更多姿的体验。

[张立行 @bylixing]

如果 Agent 时代只能记住两个词，我会选 Context 和 Loop：一个决定 Agent 这一次做得好不好，一个决定它下一次能不能做得更好。Agent 调度应该是越来越轻松的，不要让自己成为系统的瓶颈。 https://t.co/vkopXpF9e3

[张立行 @bylixing]

grok bot 确实是优雅，隐藏了很多复杂度，交付给用户的体验就比较简单。

什么 skill、harness、mcp ，不重要：）

https://t.co/MmSHoCIL7J

另外，虽然预测偏乐观，但确实大势如此。

面向 Agent 时代 的团队建设非常重要的一件事就是构建上下文，让整个组织对 agent 可读 =》 可写 =》 可闭环（loop）

[Lee Robinson @leerob]

How do always-on, proactive agents like @Bot work?

Watch my lecture at Stanford CS146S:

2:35 How we got here
6:13 What changed in the models
10:06 Inside an always-on agent
18:07 The harness
27:26 Context engineering
34:26 Where this is going

[张立行 @bylixing]

@tony56531913631 博客一直正常访问的

[张立行 @bylixing]

今年八月份羊毛价拿下了 grok+cursor 的包年会员，随赠了 X 的会员。

闲着也是闲着，干脆籍此承载一下我最近沉迷 Agent 后旺盛的表达欲：）

目前的形象是日耗 30 亿 token 的塞尔达玩家，分享下我和 Agent 协作过程中的一手实践。

[张立行 @bylixing]

@zwdroidai @chunxiangai loop 转起来很容易上并发，更多的瓶颈在于精力、cpu 算力、磁盘、token

[张立行 @bylixing]

想起个事，mac mini 放海外，如果觉得 UU 远程跨境不稳定或者 ssh 太麻烦。

可以接入这类多 Agent 调度平台，网络问题他们来搞定，各自也都有手机和电脑的客户端。

除了 cc，也可以很方便地接入 codex、grok build 一系列的 agent。

也是旧文里提到过的方案 https://t.co/l7s9Z2C1pc

stablyai/orca: Orca is the ADE for working with a fleet of parallel agents. Run any coding agent with your own subscription. Available on desktop, mobile and remote runtime.: https://t.co/ymEtewjFJg
getpaseo/paseo: Orchestrate multiple coding agents from desktop and mobile: https://t.co/kgLP1dkhT4
Cindy — The open-source AI agent, ready from day one: https://t.co/AKw05OSfKV

[张立行 @bylixing]

任何单点级别的方案都没有意义，要做就做全套，不然不如不做。 https://t.co/gdsdsrUf6M

[张立行 @bylixing]

VideoTranscript API 服务 1.0.0 服务更新：

除了支持 youtube、小宇宙、bilibili 等常见音视频平台以外，本次允许上传本地文件了。

不过所有产物原则上是公开可读的，所以请注意隐私。

自己部署的时候非必要不要暴露到公网。

如果使用 cloudflare tunnel 的形式暴露到公网，那免费套餐的最大文件限制是 100 MB，此时建议基于 tailscale 等环境上传大文件。

https://t.co/Qj2LoSs9Ho
（惭愧，开源这么久甚至都没开 releases 版本号）

[Jerry Liu @jerryjliu0]

These days, you can pretty much solve any task by defining an eval and hillclimbing over it, instead of directly defining the deterministic/agentic workflow to solve it.

The data provider companies' entire job is to define evals for all economic activity to make the frontier models capable of doing anything.

Your job then becomes pointing frontier intelligence in the right direction - defining what to solve, and how to measure what good looks like.

Agent application interfaces will evolve to capture this. The most complex processes will still need some sort of explicit workflow builder interface, but most tasks can be compressed into goals and eval instructions.

[张立行 @bylixing]

这就是 gemini 语言能力还行的原因😂

[Patrick C Toulme @PatrickToulme]

I was at an AI dinner last night in NYC.
Some interesting topics discussed

1. Performance engineers now spend most of their time managing agents and RSI loops. Most perf opts at major inference providers (together, fireworks) are found by agents.
2. RLing too much on code and math has negative effects on your model especially it’s language.
3. People were impressed by @bot and the overall acquisition of @cursor_ai by @spacexai. xAI is a legitimate frontier lab right now.
4. Most people have agents in loop writing, reviewing, and submitting PRs 24/7
5. Most agreed code review is not super productive anymore
6. Consensus was Fable 5.5 will be stunning but OpenAI will beat it in December with Bel

[张立行 @bylixing]

深有同感。

Loop 搭起来了之后很容易上并发，然后瓶颈开始在其他地方出现：

- 本机 cpu 连续打满=》完整测试去 ci 机器上跑
- 本机 worktree 动辄数百个，磁盘爆满=》建立 worktree 全生命周期的创建销毁机制
- ci 排队时间过长=》加机器、优化 ci 执行策略，分级复用结果
- token 不够用 =》 分层管理，执行层全部用牛马模型；持续优化上下文，降低单任务执行成本。

https://t.co/h1euxEINgd

[张立行 @bylixing]

如果 Agent 时代只能记住两个词，我会选 Context 和 Loop：一个决定 Agent 这一次做得好不好，一个决定它下一次能不能做得更好。Agent 调度应该是越来越轻松的，不要让自己成为系统的瓶颈。 https://t.co/vkopXpF9e3

[张立行 @bylixing]

软件工程一直都是一个管理复杂度的科学。面向 Agent 的编程和面向 Python 编程和面向 C 编程，某种程度上它只是语言的抽象程度的差异，也就是所谓的高级语言和低级语言。

在这个背景下，好的工程习惯还是有非常大的价值，所谓牛逼的模型，一个点是在于它内化了很多好的工程习惯。

人类肯定是有人月神话的，Agent 应该也有，但是二者所需要的数量级不在一个层次上。这个更多的还是要考验整个的 Agent 调度能力了

[lauren @poteto]

on a related note, there's been some chatter recently about software engineering skills (eg the ruby/rust/elixir drama). i'm a bit torn because while i do feel like they're software engineering principles and patterns are still important, at the same time i wonder if there will come a point in time when they just won't matter anymore. not that they're no longer useful, but that your own intuition can fill in the gaps as you use agents in an intuitive sense rather than knowing explicitly what they're doing

if that doesn't make sense, the analogy i have is the first iphone. 20 years ago you had to be a professional with expensive equipment, now just with a phone and some creativity you can film content that's compelling to a new audience. i don't know if there's a term for this but i want to call it "skill compression". where what used to be gatekept and prohibitive to experts becomes accessible to a new audience of creators with a whole new meta. the details have sort of faded away and now the craft is not in knowing the equipment and how to use it, but how you use it creatively

we're going to see a whole new generation of builders who've never seen or written a line of code build incredible things

[张立行 @bylixing]

本质上最稀缺的只有人的注意力了

[Calvin Grunewald @CalvinGrunewald]

There were some really good comments on this post! The ones closest to my mental model all focused on the role people still need to play when working with LLMs. There were also some good thoughts on how to define cost/impact.

So here is how I think about prioritization with agents. Instead of the prioritization orienting around features or things to build, the prioritization framework must orient around the new scarce resource: my attention.

This is how I work now. In a world where I can build anything, I need to be super intentional with my attention. And so given a particular item of work, I think about how much attention it needs and then choose the right tool for the job. More on the tools later!

The below diagram calls out this idea of the loop. I mean this in the abstract sense - a software development lifecycle process - with observable points. When a Human is the loop, you are managing the SDLC yourself. When you have a human around the loop, you are monitoring those observable points of the SDLC yourself. When an agent is around the loop, you've built a harness or some type of adversarial review system to monitor the points of the SDLC. And finally, you can of course just ask the agent to do it all (there is no SDLC - it's just the agent).

I like this framework because it flips the focus on how to optimize workflows - you go from optimizing how you build to what you pay attention to.

[离离子 @Ion_Mio_]

可能很多人还不知道，New API 的格式转换已拆成一个独立的 Go 包 RelayKit，支持 OpenAI Chat/Responses、Claude、Gemini 的转换，AstrLink 用的也是这套转换器。
我们想为社区多做些基础建设，让大家少重复造轮子，多花时间做自己的产品。
https://t.co/yqUbtedX2z

[张立行 @bylixing]

从 0-1 和从 1-100 面临的问题是截然不同的。

Agent 作为杠杆会让原来只有大公司面临的问题在中小团队乃至个人身上出现。

当你想要极致放大 Agent 的价值，那多 Agent 的调度管理-产出可信输出进生产，这件事就必然绕不过去。

因为最稀缺的永远是人的注意力。

[lauren @poteto]

volume does matter
look, i get it. if you were an engineer before agents came along, you're rightfully skeptical about lines of code and number of PRs being used as a metric for any kind of productivity. and i agree! or at least, i used to. for ease of explanation i will talk about this from the perspective as being an engineer on a large team of at least 50.

volume didn't matter before because we focused on impact. you could have high impact even with a few PRs: for example you could make a one line config change and save the company millions of dollars.

but before agents, the times where i saw volume come into the conversation were for the outliers. on the negative end you might have someone who is struggling to perform, on the other end you might have someone who was a "coding machine" archetype. a coding machine was someone who could just lock themselves in a room and emerge with a stack of PRs that solve a huge number of problems. so volume did matter, but only at the tail ends of the distribution, and combined with impact.

if you've ever worked at a large tech company before, you'll get what i mean. the thing about agents is that big tech co problems are now small-medium co problems as well.

how do agents change this calculus? well, everyone now has the ability to become a coding machine. you have incredibly capable frontier models that can write code better than any of us, at a rate much faster than we can. in this world, if you're still producing the same number of PRs as before, why wouldn't you pause to wonder why you're not being more productive? alien intelligence is here, and yet you can't outship a human coding machine?

this brings me back to why i keep talking about trust (https://t.co/7fzYZV8LGH). if you haven't put in the work to trust your agent's output, it is very difficult to scale up your productivity.

and yes, this approach does require more tokens, but i think about this in terms of cost per intelligence. before agents, this cost was very high - you needed to hire many software engineers with high 6 figure salaries to do the work. tokens are still expensive today, but are likely to get cheaper over time (https://t.co/gTBZx7pl14) factoring in the cost of intelligence. what looks exorbitant today will likely be affordable in 6 moths to a year. a single engineer with agents can do the work of tens if not hundreds of engineers, at a fraction of the cost.

the reality that i don't think has hit yet is that job of the software engineer has truly changed. our job is not merely to produce software any longer (well tbh it was never only about the code, but bear with me for the sake of the explanation), but the machine that writes the software. this idea of a software factory, or a Michelin kitchen as i like to call it, is still a topic of research. and it's the type of stuff i like to share, not to flex, but to show you what's possible when you put a lot of rigor into using agents at scale

[张立行 @bylixing]

@xiaomovps 阶跃算力不太足，自己的 coding plan 已经限购了。

这个 free 的算力就更难保证了

[张立行 @bylixing]

@xiaomovps 真实体验...

[张立行 @bylixing]

@xiaomovps 我意思是你跟它 hi 一声可以，但是过程中经常 429。。。所以，基本不可用于真实环境。
比之前的 space bunny 稳定性差太多了。

[张立行 @bylixing]

Agent 时代，大部分软件工程里的目标如果能被拆解成量化的、完全等价的函数，某种意义上可以说，被解决只是时间问题。

[张立行 @bylixing]

虽然最近的硬件成本相较于一两年前很高，但是最近我还是在买内存（d4），买 CPU，买硬盘。

两点原因：

- Agent 进企业、Personal Agent 这两块都会大量抢占产能，目前都是刚刚开了个头而已。

- Agent 并发带来的工程速度直接会让 Dev、 CI/CD 负载十倍增长，进而导致 cpu 成为开发过程中的卡点。

目前消费级 cpu 的价格还没有普遍上涨，如果团队长期有需要，是值得提前入手的。

---
AI:AM: A Level We Shouldn't Pass? Notes from The Curve + Tokens vs. Salaries & Is SaaS Cooked?
原始地址：https://t.co/S50eBSPcuF

总结和校对：https://t.co/3R91jYpEtz

[张立行 @bylixing]

大部分的 coding 场景下，agent 做得会比人类更好，剩下的也只是时间问题。这块具体看 OA 两家的分享，代码 80% 都是 Agent 写了（完善的基建、无限免费的高智力 token、极高的人才密度）

安全和规模扩展性的难点在于这块有太多水下信息都没上网，原则上给 agent 无限的时间 loop 迭代我相信能做好，但是这个出错验证的成本太高了，不如直接花钱买有经验的人。