You’ve probably been obsessing over the latest LLM benchmark. GPT-5’s score. Claude’s new reasoning trick. The model that can write poetry in iambic pentameter.
I was too. Until I saw WorkBuddy’s numbers: 13 million daily active users, 20 million monthly. In the efficiency AI category, it’s number one. And it’s not because of the model.
I spent a day tearing it apart. Three modes, full architecture, cross-verified with Tencent’s official docs. What I found changed how I think about building AI products.
The competitive advantage in AI is shifting from model capability to workflow embedding depth. If you’re still chasing better models, you’re fighting the wrong war.
The Three Layers of Dependency
Open WorkBuddy and you see a chat box. But it’s not a chat box. It’s a three-layer workspace that maps to user maturity:
- Assistant mode: One command, one task. Write an email, fetch data, summarize a doc. Feels like ChatGPT but it’s embedded in your workflow—can operate local files, connect to Feishu and WeCom. No copy-paste.
- Project mode: This is the killer. You say “write a PRD for an agent product.” It auto-plans steps, executes, and then enters a review phase. More on that review later.
- Automation mode: Set it and forget it. Daily AI industry report pushed to your WeChat. Zero human intervention.
Most AI products only do the first layer. Users come, chat, leave. Nothing sticks. WorkBuddy’s three layers create a progression: try it → trust it → delegate it. Each layer embeds deeper into your workflow and deposits assets you can’t easily take elsewhere.
The Architecture That Breaks the Mold
Here’s the part that made me rethink everything. WorkBuddy’s core isn’t a master-slave agent architecture. It’s a router + Skill marketplace.
Think of traditional multi-agent like a company: CEO agent delegates to department agents. WorkBuddy is a shopping mall: a concierge (router) points you to the right store among 70,000+ Skills.
Every Skill is a self-contained mini-agent with its own routing, tools, and execution logic. New Skills are added by describing their intent in the description field—like SEO keywords. The router reads them, matches them, and dispatches. No core code changes needed.
This is the opposite of the “build one super-agent” approach that everyone’s chasing. It’s modular, scalable, and—crucially—allows the system to grow without architectural debt.
The Truth About Multi-Agent Collaboration
WorkBuddy’s marketing says “one command, one assistant team.” Sounds like agents working in parallel from the start, right?
Wrong. And this is the twist.
In my test, I asked it to write a PRD. During the creation phase, a single agent handled everything—no splitting, no role assignment. One agent, full context, end to end. That’s smart.
But when it entered the review phase, the system spun up three expert agents: compliance, technical architecture, and interaction design. They worked in parallel, each from their own perspective. The result? 54 issues, 16 blocking. The compliance agent caught data privacy gaps. The tech agent questioned the architecture choice. The interaction agent flagged information density problems.
Multi-agent collaboration isn’t about working together—it’s about finding each other’s blind spots. The real value isn’t parallel execution; it’s cross-validation. One agent maintains context consistency for creation. Multiple agents provide diverse perspectives for review.
This is a design decision every AI content product should steal. Don’t let multiple agents write the same thing. Let one agent own the narrative, then let others stress-test it.
The Moat Nobody’s Talking About
WorkBuddy’s model strategy is brutally pragmatic: strong models for planning (Claude, Kimi), fast models for execution (GPT, GLM). No vendor lock-in. Model commoditization is a feature, not a bug—as models get cheaper, WorkBuddy gets stronger.
So where’s the moat? Not in the technology. In the assets.
After 100 days of use, a team has accumulated: Skills (custom workflows), SOPs (standardized processes), connector configurations (Feishu, WeCom, databases). These are organizational assets. They’re not portable. Switching platforms means rebuilding all of it.
The real moat isn’t artificial intelligence; it’s human organizational lock-in. This isn’t user stickiness—it’s company stickiness. Even if a single employee prefers another tool, the organization’s migration cost is too high.
Compare with Coze: One builds agents for developers (POC playground). The other builds products for office workers (daily embedding). Same technology stack, completely different business model. One captures developer attention. The other captures organizational switching costs.
What You Can Steal
Three decisions you can apply to your own AI product:
1. Architecture: Router + Modularity
Not every product needs a Skill marketplace. But every AI product needs a routing layer and modular components. Think of it like a hospital: you need a reception desk (router) and standardized departments (modules). You don’t need outside doctors coming in until you’re big enough.
2. Product: A Path to Deeper Use
Assistant → Project → Automation. Each step builds trust and deposits assets. Ask yourself: after 100 days, what does your user have that they can’t take away? If the answer is nothing, you have no moat.
3. Business: Choose Your Game
Are you building an agent for people to use or an agent for people to build? Those are two different products, two different growth paths, two different business models. Don’t try to be both.
WorkBuddy isn’t the best AI product. But it’s the furthest along the path of workflow embedding. And that’s the path that wins.
So ask yourself the three questions that matter more than your next model choice:
- What assets do users build in your product that they can’t take with them?
- Does your product have a path to deeper use, or is it a one-trick chat?
- Are you building a tool or a way of working?
The answers will determine whether your product is a commodity or a lock-in.
FAQ
Q: Why should I care about WorkBuddy if I'm not building a Chinese enterprise product?
A: The principles apply universally. The shift from model capability to workflow embedding depth is happening in every AI product category. The specific three-layer design and routing architecture are patterns you can steal regardless of your market.
Q: Isn't the multi-agent being single-agent in creation just a limitation of current models?
A: No, it's a deliberate design choice. Context consistency is critical for quality. Multiple agents writing different parts of the same document inevitably create coherence issues. The real value of multi-agent is cross-validation, not co-creation—a lesson that applies to any content generation system.
Q: What if my product is a small tool used by individuals, not teams?
A: The moat is smaller for individual tools, but the same principle applies. If your tool lets users save preferences, templates, or style guides that improve over time, they're building assets. The question is whether those assets are portable. Make them non-portable (or costly to export) and you create stickiness.