Looking at the WorkBuddy roles on Tencent Careers, I have a rough idea where it's headed next

Last night I was about to go to sleep, but my hand slipped and I opened Tencent's recruiting page, scrolling all the way down. There's a long list of WorkBuddy-related positions: enterprise agent full-stack development, open ecosystem senior development, backend development, mobile senior development, client development, in both Shenzhen and Shanghai, some asking for 10+ years of experience, some 3-5 years, with salaries listed at 50-80K · 15 months' pay. After reading I was dead tired, but I sat up straight.

An office tool hiring simultaneously for backend, client, mobile, and open ecosystem—this is a lane change, not adding features.

The JD has a very blunt section: Managed Agents product R&D, covering Agent usage, orchestration, execution, governance, evaluation, and also linking with WorkBuddy / CodeBuddy for enterprise managed agents. Another section talks about enterprise-exclusive deployment architecture, multi-tenant isolation, model service interfaces, data interaction, compliance and isolation.

These words feel pretty distant from an ordinary user. But after using WorkBuddy for over two months, the more I use it, the more I feel it's long since run into these words.

When I first started using it in early August, like most people, I opened it and just let it work. Transcribing meetings, organizing visitor records, writing parent notification templates. The capability is sufficient, really sufficient. The problem is when it touches my files.

The counseling room's files are messy. Each client gets a folder, containing intake records, brief notes from each session, and occasionally assessment scales. These files have strict boundaries between them—A's stuff can't appear in B's context. At first I didn't care about this, dumped everything in one big directory, and let it find things itself. As a result, once it mashed two clients' notes together to make a summary for me, and even got the names wrong. My back went cold at that moment.

Later I changed my approach. First create directories, one folder per client, then write a SKILL.md in the root directory, clearly stating what it can touch, what it can't, and where output goes. I was especially verbose in the forbidden-zone section, like "must not read across folders," "must not modify original records, only create new files." After the change it was much smoother.

In September I wrote a post saying that only after you install WorkBuddy into your own folder and set permissions is deployment complete; just downloading and using it is only research. Looking back now, I said that too early, but the direction wasn't wrong. That line in the JD about "multi-tenant isolation"—for me it's "one isolated folder per client." What they need to solve is that data from hundreds of departments in an enterprise can't cross; what I need to solve is that records from dozens of clients can't cross. It's essentially the same thing, just my scale is small enough that I don't need a whole system.

So when I look at those positions, I don't feel a disconnect. When they write "governance," what appears in my head is my SKILL.md.

One thing left a deep impression. Last month I threw in a dozen invoice PDFs and phone screenshots, and asked it to categorize and summarize. Most were recognized fine; two crooked photos needed manual correction. The final exported table format was correct, saving me an evening.

But it often misses the remarks column. In my testing, about three or four out of ten have empty remarks, and I have to go through them myself.

Tools like WorkBuddy—the capability has long been sufficient; the trouble is that after it's done, you still have to verify it. In enterprise scenarios this problem is more serious, because one person can't verify it all. That's why the JD wants people for "evaluation" and "quality and performance governance." This is a real, solid need.

By the way, about CodeBuddy and WorkBuddy linking up. I guess the foundation is the same agent framework, one for people who write code, one for people who don't. This judgment may not be right, but from the job descriptions putting the two products together, it's probably inferable. If that's true, then WorkBuddy's moat lies in whether it dares to touch files on your hard drive, and whether after touching them it can clearly explain what it touched.

This is also where I think its risk is greatest. Daring to touch files is why it's stronger than a chat box, and also why I glance at the directory every time before I hit run.

Let me say something unpleasant.

The appearance of a "mobile senior development" position—my first reaction is it wants to get into phones. Rationally this is a good thing; many of my thoughts pop up on the couch, in the gaps while waiting for clients, when I only have a phone at hand. But I doubt what it can do on a phone. The permission model is already hard enough on desktop; moving it to mobile, you either cut features or cut boundaries. I lean toward it cutting boundaries.

What concerns me more is the focus. Enterprise managed, multi-tenant, compliance isolation—every time these words appear, product resources shift a bit that way. The thicker the enterprise version gets, the more easily the personal version becomes an experience version. I've seen this pattern too many times—office software, note-taking software, collaboration tools, no exceptions.

I don't like this direction, but I understand it. Enterprises pay, individuals don't. I'm just a bit wistful, because the place I need it most is exactly the place that doesn't make money: a psychological counselor, one person in one room, dozens of clients, who won't buy any enterprise version in a year.

But there's another possibility. If that enterprise governance is done solidly, it feeds back into the personal version, and users like me get a stricter permission control for free. By then I might not have to hand-write SKILL.md anymore; it'll ask me itself, "Do you want to isolate this folder?"

Which possibility is more likely, I can't say. What I'm more curious about is another thing: once it really finishes that enterprise stuff, will small users like me—one person, one counseling room—still count as its target users?