WorkBuddy Topped Two Leaderboards, but I'm Stuck on One Word in the Report: Harness

Last week I came across Sullivan's "2026 Global Desktop AI Agent Market Research Report," and Tencent WorkBuddy ranked first on both the China personal and enterprise desktop agent lists. The image was one of those standard ranking screenshots, and the comments section was full of congratulations. Before that, Omdia and IDC had also published evaluations, and WorkBuddy scored well in both. Sullivan's report is the third one.

My first reaction was to think of that file I got stuck on Wednesday night last week.

Let me give some background first. My daughter Duoduo is in second grade, and this semester the teacher asked parents to prepare a growth record for the parent-teacher meeting. I had three kinds of things scattered around. A dozen or so PDFs, including notices sent by the school and Duoduo's reading check-in records; one Excel sheet, where I recorded wrong answers and corrections myself; and dozens of screenshots on my phone, conversations where Duoduo used an AI chat tool to ask about homework. I wanted WorkBuddy to turn them into something printable, and also usable as material for my own media account.

I've been using WorkBuddy for two months. Before, I mainly used it to organize information and make PPTs, and single files basically never went wrong. This time I dumped all three kinds of things into one folder and wrote a fairly long instruction, letting it classify, extract, and merge on its own. The result got stuck at step two. It finished reading the PDFs, but when it came to aligning them with the Excel sheet, it started going back and forth, revised three rounds, and the date formats all got messed up, while OCR from the screenshots mixed in some irrelevant content.

At the time I thought it was because there were too many files. Later I split it up and redid it: first feed only the PDFs, run it and check the results, unify naming as "date plus type"; then handle the Excel separately; for the screenshots, I first cropped out irrelevant edges myself, leaving only the conversation body, then imported them. The whole process took about forty minutes, much shorter than the two hours I spent messing around the first time.

I'm telling this story because there was one sentence in the report that I think is far more important than the ranking itself. The report says Harness is the deciding factor for desktop AI agents, and it cites WorkBuddy's skill marketplace, enterprise system API and Webhook integration as examples of execution interfaces among the harness pillars.

The word Harness translates roughly as "restraint device." In AI agents, I understand it as: no matter how smart the model is, there has to be a set of things managing it, how tasks are broken down, how tools are called, how errors are recovered from, how much permission is granted, where memory is stored. In the report, the "Product Technology Leadership Index" column lists Agentic capabilities, Harness engineering, model adaptation, and extension and openness capabilities, which cover exactly these. And the "User Deployment Effectiveness Index" column, with installation and deployment, UI interaction, multi-device collaboration, workflow templates, task orchestration, scheduled dispatch, and audit compliance, are also different sides of the same thing.

The report splits "smart" and "reliable" into two columns, and then says the reliable column is the watershed.

I agree with that judgment. Because where I got stuck last Wednesday was not about whether the model was smart.

It read PDFs very well, and could even recognize dates from Duoduo's messy check-in records. What it got stuck on was that the process had no step-by-step acceptance mechanism, and that I mixed three things with hugely different formats together without it forcing me to split them first. That was my operational problem, but it's also where the product should have thought one step ahead for me.

A few days ago I wrote a post saying that WorkBuddy gets stuck when handling PDFs, Excel, and screenshots mixed together. At the time a friend replied that it might be a file naming and path issue. Looking back now, the conclusion of that post was only half right. Naming and paths do have an impact, but the main problem was that I didn't run each step separately for acceptance at the time. This way of thinking is the same thing as what the report calls harness.

WorkBuddy doesn't do badly on this point. Its skill marketplace is open, and skills like PDF reading, table organization, and document summarization can be installed and run separately, which is how I split it later. It also has scheduled dispatch and workflow templates, so in theory it can lock down things like "organize once a week." The enterprise version is said to have team memory, AI asset management, and audit compliance. I haven't used that part, so I won't comment blindly.

But there are also things I don't like, and I'll say them clearly. It is more sensitive to file naming and paths than I expected. The same PDF, placed on the desktop versus in a subfolder with a Chinese name, will produce different recognition results. This experience is not very friendly for ordinary users, especially people like us who toss files around casually. In multi-step tasks it occasionally gets too clever. I clearly said do A first and then B, but sometimes it brings B's intermediate results into A, and the output looks complete but is actually wrong as soon as you check it. My current workaround is to stop at each step and only move on after acceptance. Slow, but stable.

There is one more point, and it was my biggest takeaway this time. The report mentions zero-cost access to custom models and lists it as its "unique moat." The meaning of this capability for a personal user like me may be different from that for enterprise users. I don't really switch models much, but I know the model can be switched, and I won't be locked into one, which is a plus when choosing tools. But honestly, ordinary parents won't care about this. What they care about is whether it can get the job done in one go.

So my judgment depends on the person. If you have a single type of file, one PDF or one table, WorkBuddy is basically ready out of the box, and it's no surprise it can take first place on the personal side. If you're like me and need to process several formats mixed together, then right now it's more like a very capable subordinate that requires you to act as the foreman yourself. You have to know how to split tasks, check results, and name things, and only then can it perform. For enterprise users, the team governance and audit compliance the report mentions are where the gap really opens up, but that's already beyond my scope of use.

I've now set a rule for myself: feed only one type of file at a time, unify naming as date plus type, run each step and review it myself, and only move to the next step after confirming. I've used this rule for about a week and haven't crashed again so far. Duoduo's growth record was finally printed out, eight pages in total, and the teacher even asked me where I made it.

Over the next year, the word harness will probably move from industry reports to ordinary users' lips. Rankings will look more and more similar, and model capabilities will get closer and closer, so the remaining difference lies in who can make step-by-step acceptance and step-by-step rollback the product's default actions, without users having to figure it out themselves. My current rule was forced out exactly that way.