ZHENG XIANG | Building an AI-Native Org

10+ yrs in big tech | ex-Tencent dev lead | running a console & PC gaming network on WeChat, 2M+ users | building AI & gaming products

[ZHENG XIANG | Building an AI-Native Org @ZanneZheng]

@yigittbuilds When it's working on something hard to undo, like a production config or a repo with no backup. Read only lets it look around and propose a plan first, and you approve before anything gets written.

[ZHENG XIANG | Building an AI-Native Org @ZanneZheng]

Codex, Claude Code, OpenCode, DSH and 10+ more AI coding tools on your computer, managed from one page. When you're out, you can keep directing them from your phone and preview the web pages they build. That's OpenTandem, a partly open-source project we made.

It fixes three hassles we ran into every day.

The first is at the desk: several AI coding tools, each in its own window, plus a few Mac minis managed over SSH. Switching back and forth wears you out. OpenTandem puts multiple computers and AI agents into one page.

The second is after you leave the desk: you want to check a task's progress or add one more instruction, but you used to have to wait until you got back. Now you open OpenTandem on your phone, find the same task, and keep it moving.

The third is also after you leave: the localhost pages AI starts on your computer won't open on your phone, so you could only see the result back at your desk. OpenTandem has remote preview, so you open them right on your phone, then send your changes back to the same task.

The Android app is ready to download from our website, there are desktop apps for Mac and Windows, and a web version runs in any browser on your phone or computer. The iOS app is on the App Store outside China; the China listing is still going through registration.

We've open-sourced the Runner and CLI on GitHub, so you can read the code yourself. If you run an OPC or work in an AI-native team, or you want to use your commute and weekends to keep AI talking with you and getting work done, give it a try. Ask us anything in the comments.

Website: https://t.co/f48EVmpwSj
Open source: https://t.co/Ilfj32sgx7

[ZHENG XIANG | Building an AI-Native Org @ZanneZheng]

Besides previewing, you can answer the agent right from your phone: when it asks you something, tap an option to approve or type in other info. As for permissions, it works the same as the coding agent on your computer. You pick Full access or Read only, following the same logic as the agent you use.

[ZHENG XIANG | Building an AI-Native Org @ZanneZheng]

@kidtsang Thanks, agreed that the feedback loop matters more than raw output.

[ZHENG XIANG | Building an AI-Native Org @ZanneZheng]

@wangyan_im Right now it's the page itself, and you can ask the agent to check the console and tell you what's wrong.

[ZHENG XIANG | Building an AI-Native Org @ZanneZheng]

A quick look at a workday for my 9 AI employees in OpenTandem. Each avatar shows what that employee is up to.

The one bouncing up and down, with its name in bold and a number next to it, has just finished a task and is waiting for me to check the result.

The ones with a keyboard in front and eyes darting around are busy working.

The ones sitting quietly have nothing on their plate and are waiting for me to hand out work.

[ZHENG XIANG | Building an AI-Native Org @ZanneZheng]

@EmiliOlivi_ai Probably the final pacing. I’d hand off the tedious trimming, but deciding when a moment has landed still feels like a human call.

[ZHENG XIANG | Building an AI-Native Org @ZanneZheng]

@monymonykay The gap between read-only and write access feels like the real bottleneck. An agent can look capable in a demo; getting a team comfortable letting it act is a different bar.

[ZHENG XIANG | Building an AI-Native Org @ZanneZheng]

@SHAMALKRISHNA That price jump is easy to miss if you’re only watching the model string. Glad you called out both dates.

[ZHENG XIANG | Building an AI-Native Org @ZanneZheng]

AI is a building set for grown-ups. Children use blocks to bring their imagination to life; adults use AI to bring their ideas to life.

In an interview, Boris Cherny, who started the Claude Code project, suggested giving Claude tasks that seem a little too hard, along with a way to verify its own work. In the AI-native era, we should give our creativity that kind of room, too, without asking how much money an idea can make the moment it comes to us.

Watching my three-year-old build with blocks for three hours straight, I was amazed that such a restless little boy could become so absorbed. As a middle-aged guy, I finally started to understand why kids love Minecraft: this kind of building gives them so much freedom to express their creativity.

Then I look at the people around me. So many are using AI to make software, build their own agents and apps, and bring the ideas in their heads to life, piece by piece. The joy of making something their own way draws adults in, too. Some have set aside their old hobbies; some have even left their jobs to focus on what they’re creating.

If you want to make something with AI, try approaching it with a child’s simple curiosity: when an idea comes to you, give it a go. Pick one small idea you’ve always wanted to bring to life. Turn it into something people can actually use and try. Then find people who share that idea but can’t build it themselves. Enjoy the process of making it first, then test its commercial value.

[ZHENG XIANG | Building an AI-Native Org @ZanneZheng]

Claude Code accounts keep getting banned, yet people keep signing up. My data shows why.

$200 plan on both, maxed out Oct 15–30:

1. Usage: 11.44B vs 4.10B tokens (2.8x)
2. Time per run: 7.7 vs 11.4 min (32% less)
3. Failure rate: 1.2% vs 3.1% (62% lower)

#ClaudeCode #Codex

[ZHENG XIANG | Building an AI-Native Org @ZanneZheng]

@CarinaAyden Reliability and cost are the parts that seem easiest to underestimate until an app is actually in use. Curious which tool in the collection surprised you most.

[ZHENG XIANG | Building an AI-Native Org @ZanneZheng]

@techamericaofcl I’d draw the line at whether the result is recognizably borrowing a specific game’s identity, not whether it shares a familiar mechanic. That still leaves plenty of room for experimentation without turning a prototype into a clone.

[ZHENG XIANG | Building an AI-Native Org @ZanneZheng]

@Saadijplays That’s a neat pipeline. The handoff from research to visuals is what makes this feel like more than just another AI-written tweet.

[ZHENG XIANG | Building an AI-Native Org @ZanneZheng]

@Binyu63 I still reach for Google when I want to compare sources, but AI is becoming a useful starting point for product discovery. I’m curious how often it actually leads to a purchase.

[ZHENG XIANG | Building an AI-Native Org @ZanneZheng]

@NatashaKurldso Having to hunt for a way to ask for a quote is such an avoidable bit of friction. It’s surprising how many sites still make that the visitor’s problem.

[ZHENG XIANG | Building an AI-Native Org @ZanneZheng]

@ichkrub Welcome back! Looking forward to the AI tools you actually use—those firsthand notes are always more interesting than another generic tool list.

[ZHENG XIANG | Building an AI-Native Org @ZanneZheng]

A one-image guide to "Super-agent Design Exploration", the next big trend for agents.

The idea isn't new: https://t.co/FPIGYFgksl has practiced it for a while, as does the AI employee module I'm building in OpenTandem.

The long original, in one plain-language image.

#AIAgents

[ZHENG XIANG | Building an AI-Native Org @ZanneZheng]

One person, 30 days, 23.47B tokens. It's the best time to start up: a solo founder can run like a company.

AI runs it all, from strategy to social, on 5 computers via OpenTandem.

2,355 commits, 27 videos, 107 releases, 271K new lines.

Cheers to this era 🍻

#OPC #buildinpublic

[ZHENG XIANG | Building an AI-Native Org @ZanneZheng]

Finally: Claude is phasing out the 5-hour cap for verified Ultra (formerly Max 20x) accounts. Weekly limits stay. Getting stopped mid-task in Claude Code felt like squeezing out the last bit of toothpaste. Glad this is changing. #OPC #IndieDev