A Subjective Take on the Agent Wars Among ByteDance, Alibaba, and Tencent

"Office Agents Need a Melvin Chen Too"

After objectively evaluating the Agent wars between ByteDance, Alibaba, and Tencent, we're continuing to play 4D chess — this time from a purely subjective angle, on the paths office Agents might take.

Step one: the Agent form factor converges into a local client + chat box + a mobile app that controls your computer.

That might sound complicated. Let me put it in plainer language: copy Codex.

Whatever Codex ships, everyone ships. No thinking required.

Someone told me earlier that IDE interfaces are amazing, that Qoder targets high-productivity users, and that its data flywheel is more valuable than Qwen Office's.

That's pure nonsense. None of the programmers around me use IDEs — why would you use an IDE when a chat box is this simple?

Sure enough, Qoder's latest revamp has also switched to a chat box form:

There was also some cute investor saying maybe we should stop investing in GUI-minded software companies. My suggestion: he should be restricted to CLI-only work from now on. Codex is banned for him.

All you early-adopter geeks, stop acting superior with your "the future of Agents is CLI" and "Agents will be a user base ten times the size of humanity." That's all storytelling. If AI is going to serve the masses, it has to be GUI.

Once the product form converges, the org has to converge too. Tencent, Alibaba, and ByteDance all need to consolidate into one product, one entry point. Xianyu has covered this already so I won't repeat it — I'll just make one point:

Can you imagine, during the Spring Festival ad war, ByteDance pouring money into Doubao Work while Alibaba promotes Qoder and Qwen Office at the same time? Which Agent am I supposed to use to order milk tea and grab red envelopes?

Unless, of course, you can claim both — then no merger needed.

All I can say is, I thought Qoder launching Qoder Work meant accepting the org's arrangement and merging into Qwen Office. Turns out it was a sacrifice-the-pawn-save-the-king, all-or-nothing move.

Qwen launching late and getting flattened by Doubao — fine, there's an excuse. But Qwen Office launched first, had a model advantage, and still got overtaken by Doubao? That's purely an org problem.

ByteDance next door, meanwhile, looks like it could shut down both Coze and Trae any minute. Disclaimer: we absolutely do not encourage this kind of behavior. It's a bit terrifying — a full-on Qin-dynasty company.

The route for our Greater Khingan Mountains road trip was made with Qwen Office.

Step two: hardware-software integration.

One thing I've always believed: hardware (at this stage, the recording card) is a crucial entry point for office Agents.

Think about it: when each company's sales team sells its office Agent, if they bundle in a recording card, wouldn't renewal rates double? It also raises the cost for competitors trying to poach your customers.

And for the organization, giving everyone a recording card means more frequent use of the bundled office Agent, plus more complete context accumulation.

How does that context eventually get used? No idea. At minimum, if the company goes under, you can sell the data to model labs and fund everyone's severance.

I already argued yesterday that the recording card isn't a standalone market in China — it's an entry point for office Agents. Plaud flopping in the Chinese market was inevitable.

Meanwhile, DingTalk and Lark selling recording cards at near-cost is pure infrastructure-building. We're now in the recording-card infrastructure era.

Of course, the branding should be unified sooner rather than later. My suggestion: rename the Lark recording bean to the Doubao recording bean, and the DingTalk recording card to the Qwen recording card.

To maximize user data and close the loop on user context, the hardware must be self-owned and self-developed.

So let me say it again — I sincerely recommend WorkBuddy acquire Plaud, the sooner the better. WorkBuddy's open ecosystem conference left zero impression on me. "Every piece of hardware can integrate" is basically equivalent to nothing integrating.

Turning it into QQ Show has real potential

As for whether anything besides recording cards — glasses, watches, earbuds, Vibe Coding keyboards, even the Doubao phone — could be an entry point:

All I can say is, everything has a chance. It all depends on our Shenzhen hardware family discovering PMF for the big companies to copy.

Step three: office Agent + IM (instant messaging) + accumulated enterprise context, entering the era of AI collaborative work.

The VC market's Q2 consensus was AI Workspace — the idea being human-AI collaboration, with the ambition of killing DingTalk and Lark.

Coze and MuleRun clearly both realized this feature is a must-have and rushed out similar updates. But they both lack enterprise context, and their orgs have other plans for them.

Among the startups, Raft is probably doing it best — but it's still a niche hobby.

Now things are different. Lark and DingTalk have become feedstock for office Agents. At this stage their main function is funneling traffic to Agent products, but once everyone gets used to working through Agents, two people using an Agent to do the same task becomes very natural.

Xianyu and I, for example, now take turns in Raft having the same Agent filter event sign-up sheets.

It's just that right now, neither humans nor Agents have enough context, and few people can adapt to inserting an Agent into human-to-human collaboration.

Once collaboration happens, you naturally need IM to communicate — and you need the Agent to read conversations, group chats, and documents to assist with work.

So yeah, if these office Agents really want to disrupt Lark and DingTalk and enter the enterprise collaboration market, they need a Raft.

My meaning is obvious enough — let's see which of Tencent, Alibaba, and ByteDance snaps up Raft first.

Side note: once everyone has their own office Agent, Baidu Search — already on its way to being swallowed by Xiaohongshu — is really finished.

Of course, all three of these steps are pure speculation, and they could all happen simultaneously. Qwen Office recently shipped a multiplayer workspace — stitching things together early is fine, it's just hard to reach a true human-AI collaborative form right now.

Feature-wise, every office Agent will end up identical. Even a roadside product like Baidu Dazi — there's hardly a feature anyone can't build. Here's my prediction: office Agents, like video Agents, will enter the Great Homage Era.

As for whether to let users pick their own models, each company has its own thinking.

Doubao Work, for instance, looks committed to one path to the end — believing users ultimately won't need to choose models, and resolutely staying closed-source. That's a company trait.

For WorkBuddy, opening up model selection is actually the easy call. Hunyuan is a second-tier model anyway, so WorkBuddy loses nothing by not using it — it can even spin a narrative like "giants battle for AI office, no longer just competing on models."

Because at this stage, users' office needs can be fully met by each company's Flash-tier models. New model releases mostly serve as user acquisition. For truly high-difficulty tasks, people go to the Silicon Valley twins: Claude Code and Codex.

Also, I think current pricing is unreasonable. WPS has already figured out that Chinese users spend under 20 yuan a month on office software, but the Big Three are all priced around 60 yuan.

Flash models could absolutely bring prices down and let real C-end users in. It's just that nobody wants to fire the first shot in a price war.

What can I say — the office Agent market needs a Melvin Chen to bring prices down for everyone!

Come to think of it, office Agents could totally run on Melvin Chen's gym-membership logic: most people never use up their quota, and the rest is pure profit!

One last bulletin: just yesterday, WorkBuddy finally figured it out and brought Peace Elite into WorkBuddy. Honestly, if they'd started handing out Q Coins in WorkBuddy earlier, who would use Doubao Work during back-to-school season?

Hoping for an integration with TiMi Studio soon — I want WorkBuddy to send me skins 🫡

(This article's cover was generated by ChatGPT; the writing is 100% human.)

This Friday at 8 PM, we'll be live with MuleRun's former head of product Chengrui Fu to roast the Big Tech Agent wars. Come watch.

⬇️

Subscribe to our Substack: funeralai.substack.com