WorkBuddy’s “Human‑Wave” Trap: Is Its Open Platform Empowerment or a Cop‑Out?

On September 2, Tencent’s WorkBuddy threw open its Agent foundation to more than 100 partners. Hardware makers like Plaud and Rokid jumped in. Enterprise apps like Tongdaxin and GF Securities followed. Nine co‑branded devices launched the same day. The press release spoke of an “operating system for the Agent era.” The message was clear: Tencent is going all in on AI.

But scale isn’t the real story. The real question is one of design philosophy: when a platform forces users to manually route its own tools, is that empowerment — or just a clever way to offload cognitive cost?

Three Layers, One Headache

Under the hood, WorkBuddy’s architecture is beautifully modular. It splits into three clean silos:

• Expert – a persona you can summon (e.g., “Sales Analyst Expert”).

• Skill – a downloadable capability (e.g., “PPT Generator”).

• Connector – a link to external data sources (e.g., Google Drive, Slack).

For an engineer, this is elegance. For a user, it’s a maze.

Imagine you’re a product manager. You walk into WorkBuddy with one sentence: “Analyse why our sales dropped this quarter and build a deck for the CEO.” What happens? Instead of the AI figuring out the steps, you’re presented with a decision tree:

• Which Expert – data or business? Or assemble a panel?

• Is the Excel Skill installed? The PPT Skill – but there are six versions of that.

• Which Connector – Google Drive or WeChat Work?

• Should I turn on Deep Research? Will two Skills conflict?

The AI was supposed to orchestrate. Instead, you become the conductor — and you weren’t trained for this.

The Expert Explosion

WorkBuddy defines an Expert as a combination of three variables: role, methodology, and tool boundary. Change any one of those, and you’ve got a brand‑new Expert. That’s how you end up with a gallery that looks like this:

PPT Expert, Presentation Expert, Slide Expert, Consulting Deck Expert, Business PPT Expert, McKinsey‑Style PPT Expert, Pitch Deck Expert, PPT Beautification Skill, PPT Generator Skill, PowerPoint Skill…

Try telling them apart. You can’t.

In the old app‑store world, quantity was a virtue. WeChat and Alipay have clear, non‑overlapping jobs. But Agent Skills are not apps — they are fragments of capability. Their boundaries are inherently fuzzy. Two hundred Experts may only represent twenty underlying functions. When users can’t see the difference, they don’t experiment — they just close the tab.

This is not a UI problem. It’s a worldview problem. WorkBuddy treats AI like a labour market — you hire the right worker for the right task. But users don’t want to be HR managers. They want results.

The Quiet Contrast: Codex

OpenAI’s Codex (the engine behind ChatGPT’s advanced data analysis) takes the opposite approach. You don’t pick a “Data Scientist Expert.” You simply say, “Analyse this data and visualise it.” Behind the scenes, Codex decides which internal Skills to invoke — and it has over 100 of them. But the user never sees that menu. They only see the output.

Codex compresses complexity. WorkBuddy exposes it.

This is not a technical advantage for one or the other. It is a philosophical fork. One path says: “The AI should be intelligent enough to figure out its own tooling.” The other says: “We’ll make the tooling transparent, and the user can decide.” The latter sounds democratic, but in practice it’s exhausting. Every decision is a tax on attention.

Hardware: The Smart Move, but Not the Cure

WorkBuddy’s push into hardware — smart glasses, recording cards, microphones — is strategically brilliant. It gives the platform physical “data tentacles” in the real world. That’s something Codex can’t easily replicate. It’s a moat.

But hardware doesn’t fix the core design flaw. If the software experience is still a jungle of dropdowns and toggle switches, adding more devices only multiplies the confusion. A smart microphone that feeds into WorkBuddy still requires you to pick the right Expert to process the recording. The hardware is a data pipe; the cognitive bottleneck remains.

The Ghost of Yingyongbao

Tencent has played this game before. Yingyongbao, its Android app store, was a channel‑dominance play — aggregate as many apps as possible, and users will come for the selection. It worked, but it never became the defining platform of the mobile era.

Now WorkBuddy is trying the same recipe in AI: recruit partners, open the platform, and bet on network effects. But AI is not mobile. In mobile, apps were complete products; users chose one and stuck with it. In AI, Skills are composable fragments — they need to be stitched together intelligently. Leaving that stitching to the user is not an ecosystem; it’s a design debt.

What Agent OS Should Look Like

In the long run, concepts like Expert, Skill, Connector, MCP, and Sub‑agent are infrastructure — like threads, drivers, and DLLs in an operating system. They are critically important, but ordinary users should never see them. They are for the AI to understand, not for the human to manage.

When you open a web browser, you don’t configure TCP/IP settings. When you use a search engine, you don’t pick which index to query. Great technology disappears.

WorkBuddy’s biggest mistake is making these internal components visible and actionable. That’s a sign of insufficient confidence in the underlying model — a belief that human oversight is still needed to patch the gaps. But in 2026, with models advancing at breakneck speed, that patch is getting thinner every quarter. What worked as a temporary crutch is becoming a permanent burden.

The Final Reckoning

WorkBuddy’s goal is understandable: reclaim the platform crown that Yingyongbao never quite seized, this time in the AI era. But channels and partnerships are amplifiers, not compasses. They can scale a good idea, but they can’t fix a flawed one.

If WorkBuddy continues to treat its users as unpaid AI‑routers, it will fall into a classic trap: a huge ecosystem that nobody knows how to navigate. The launch will be remembered not as a watershed, but as a spectacular act of self‑consolation — a grand party where the guests showed up, but nobody knew what to do once they got inside.

Confidence in model capability is the only brush that draws a real roadmap. Resources follow; they do not lead.