
WorkBuddy Scored Eight Perfect Marks, So I Opened QoderWork to Compare Again — It Won in One Area
The most expensive thing in my shop is that spreadsheet holding three years of orders.
Last Wednesday around 11 p.m., I was replying to customer service while having it organize the "To Archive" folder by month. The next morning I opened it and found a batch of files missing. It had directly deleted several old tables whose names looked like duplicates. My hands went cold. Three years of purchase prices and refund records were all in those tables.
Later I recovered them from WorkBuddy's own recycle bin. By default it sends deleted things to the recycle bin, and it keeps a backup before making changes. I sat there for two minutes before it occurred to me to check its operation log, to see which files it touched, at what time, and what state they were in before—it listed them one by one pretty clearly.
After this I seriously looked at its security documentation. Coincidentally, around those days I saw a news item: on November 12, the Saisheng Lab under the National Industrial Information Security Development Research Center released an "Office Agent Product Application Security Evaluation Research Report," doing a horizontal hands-on test of several mainstream domestic office agents. WorkBuddy got full marks in eight of ten security dimensions, passing 66 of 68 test cases.
The report mentioned that in the past three months, over 300 problematic skills and experts were blocked before listing.
My first reaction to that number was: so people really do stuff bad things in there.
I've used Skill for two weeks, installed about six or seven, and never thought about whether I should check what permissions it wants before installing. Later I learned that WorkBuddy automatically scans third-party Skills before installing, and only lets you enable them after scanning. In my testing, when installing a table-processing Skill of unknown origin, it popped up a prompt saying this skill would read clipboard contents and asked if I was sure. I clicked cancel. In the past, I'd most likely have just clicked next.
So what I want to do in this post is simple: put WorkBuddy and another one I have, QoderWork, side by side and compare item by item. What I'm comparing is: in a small shop like mine where one person does the work of five, which one can really make you feel safe handing off the work.
On security, WorkBuddy wins, and wins fairly clearly—I'll put this first. But QoderWork genuinely wins on one point: speed. I'll get to that later.
First, privilege escalation. I asked WorkBuddy to change one notes column in the "November Orders" table, and it only touched that column. Once I deliberately spoke vaguely, saying to clear out the duplicates in the order table, and it didn't act directly—it first listed a checklist asking which table, and what counts as duplicate. On QoderWork I tried a similar phrasing, and it just started running, ran for about twenty seconds, and touched two tables. Nothing went wrong, but my heart skipped a beat. This difference is a difference in design orientation: one asks first by default, one acts first by default.
Second, Skill origin. WorkBuddy requires every skill to pass a security check before listing—code and behavior are scanned, developer identity is verified, and it keeps checking after listing. I can't verify this from the user side, but at least it gives me a prompt when installing. On QoderWork, when I installed, I didn't see a similar security scanning step in the interface—though maybe I just didn't find the right place, so I won't draw a firm conclusion.
Third, desensitization. This is what I care about most. Every day I export customer lists to send to couriers, full of phone numbers and detailed addresses. When WorkBuddy exports, it masks the middle four digits of phone numbers, and addresses only keep down to the district. At first I found it annoying, but then I thought: these tables I send to group chats and to dropshippers—every extra hand is extra risk. It recognizes locally, doesn't upload, and the report mentions this too: input box, tool returns, and data egress—all three channels are handled. On QoderWork I didn't find a corresponding switch; export is export as-is.
Fourth, audit trail. WorkBuddy keeps a record after finishing work; I checked once and could see which skill it called, which files it read, what it generated. This feature is useless day to day, but it's a lifesaver when something goes wrong. Like those deleted files earlier—without records I could only stare helplessly.
Now it's time to talk about where QoderWork wins, otherwise it looks like I'm biased.
Every morning I need to process about 120 customer service messages, sorting them into four categories: refund, exchange, urge shipment, ask size. For the same batch of messages, QoderWork finishes in about two and a half minutes, WorkBuddy takes about four. A minute and a half difference doesn't sound like much, but my mornings are tight—while waiting for water to boil, a minute and a half is enough to stick a stack of shipping labels.
But accuracy is the reverse. In my testing, WorkBuddy misclassified 7 out of 120, mostly the kind where urge shipment and ask size are mixed together—customer asks two things in one sentence. QoderWork misclassified over 20, and the errors were more scattered, with a few going straight into Other. Misclassification means redoing it, and redoing takes four minutes—do the math and WorkBuddy actually saves time.
This is also my most practical takeaway from these two months: the real time-saver is not having to redo it.
By the way, let me mention the pitfalls I hit. At first I wrote the Skill's classification rules too coarsely—just the four words refund, exchange, urge shipment, size—and over a hundred messages all crowded into Other. Later I changed it to three to five real example sentences under each category, plus priority labels, e.g., if a sentence has both refund and size, prioritize refund. After the change the accuracy clearly went up. Another point: don't be greedy and have it do several things at once. I once tried having it classify while generating a summary table, and the fields all got misaligned; the cost of rerunning was much higher than running separately. Serial beats parallel—this is a lesson I paid tuition for.
Back to security. At first I thought these evaluations had little to do with a small shop owner like me—that's for enterprises. But on reflection, no. As an individual user, I have no IT department, no one to review skills for me, and no one to back me up if something goes wrong. WorkBuddy's four lines of defense—pre-identification, in-process interception, post-event recovery, full-process audit trail—are essentially stuffing the work that IT should do into the product to do for me. The recycle bin and backup—I've actually used both. The sandbox and privilege-escalation interception—I hope I never need them in my life, but it's reassuring that they're there.
As for the enterprise edition stuff—TC-AICC trusted computing, TLS 1.2+ encryption, managing shadow AI through Tencent iOA—my small shop doesn't need it, and I don't plan to pay extra for it. On November 9 I wrote a post inferring the enterprise edition from Tencent's job postings, saying that the stability improvements driven by the enterprise edition might benefit the personal edition too. Looking now, that judgment still holds up—the security investment ultimately gets spread across the whole product.
Finally, a prediction—if I'm wrong, come back and curse me this time next year.
I think by the first half of 2027, security evaluations like this will go from a bonus to a ticket to entry. Right now the leaderboards are still comparing who has more features, who runs faster; once a few real incidents happen, the first thing procurement people will ask is "did it pass the evaluation." By then, the top few can still keep up, but small vendors basically have no chance, because continuous security checks require people to maintain—over 300 problematic skills blocked before listing means a team watching behind the scenes. For individual users this is good in the short term—products will be more stable; but I estimate the free tier will shrink, because security has a cost, and someone has to pay it. I'm mentally prepared: next year it's either a price increase or feature tiering. When that day comes, I'll still renew—after all, my three-year order table was only lost once, and I don't want a second time.
Physix Frontier