Three months with WorkBuddy: the 'local execution' standard doesn't quite fit construction sites

I got back from the site a bit after eight in the evening, tossed my hard hat on the desk, and scrolled into a comparison article that lined up AiPy, WorkBuddy, LobsterAI, OpenClaw, and Cherry Studio, with the core criterion being "can it run code locally, operate on files, and deliver results." My first reaction after reading it was that the criterion itself isn't wrong, but for someone like me working on a construction site, it's a bit off-target.

I've used WorkBuddy for three months. Putting together weekly reports, organizing materials, processing submission-and-review forms — these are things I do every day. If you ask me whether it counts as a "local-execution Agent," I have to ask back first: when you say execution, do you mean running code, or delivering something?

Last Wednesday evening, the project department needed to submit a batch of submission-and-review forms for sub-item works. Seventeen PDFs — some scanned after the supervisor signed, some screenshots sent over WeChat by the construction crew, and two shot on a phone, crooked, with half a hand still in the frame. The filenames were even messier — half of them were things like "scan001."

In the past, this kind of work meant me going through them one by one, sorting by sub-item, renaming, then making a summary table. A whole evening gone.

This time I threw them all into one folder and said one sentence to WorkBuddy: sort by sub-item work, rename to the format "sub-item-date-unit," and finally produce a summary table listing each one's date, unit, and amount.

It took about forty minutes. I didn't touch it in between — went and took a shower, came back, and the folder was already organized and the summary table was out. It misread the unit on two of the crooked photos, and I fixed those by hand. A few entries in the remarks column were missing, and I filled those in too.

This is what I understand execution to be. I care more about whether it can read through my pile of junk files and put together something I can hand in directly; whether it can run a bit of Python on my computer is a different matter.

The comparison article said AiPy takes the Python-Use route, able to generate and execute code, stringing task understanding, plan decomposition, execution operations, and result feedback into one flow. I accept that — technically it's genuinely stronger. But I'm looking at the other end: the deliverable. No matter how beautifully your code runs, if what you hand me in the end is a table in the wrong format, I still have to redo it.

WorkBuddy's advantage on this is that it doesn't agonize over how to do it — it fixates on what the result should look like. I want a summary table, it gives me a summary table; format, fields, sorting — I say it once and it basically matches. Of course that doesn't mean it's better than anyone. I haven't used AiPy deeply, so I shouldn't talk nonsense. I can only say that from my testing, WorkBuddy is good enough at reading junk files.

But there are a few things about it I'm genuinely not satisfied with.

First, give it two things to do at once and it gets confused easily. I wrote a post about this in early November, saying WorkBuddy goes haywire when I give it two tasks at once. Since then I've learned my lesson — one thing at a time, and it's actually faster and more accurate. If you ask it to both sort and summarize, it tends to drop the middle step. Split it into two conversations — sort first, take a look, then summarize — much more stable. This flaw isn't small: once the work gets complex you have to babysit it and break it into steps, which means part of the time it saves I end up paying back.

Second, I don't dare throw sensitive files into it. Contracts, bid quotations, the client's internal stuff — I handle all of these by hand. No matter how convenient WorkBuddy is, I still have to check what the file is. That's not its fault; it's my own bottom line. I just picked up an ima knowledge base and am still trying it out — one more route for organizing materials — but I still don't touch the confidential categories.

Third, the execution process isn't very transparent. I only find out how it did something after it's done. If it misunderstood a step in the middle, I have to wait for the result to come out to notice. On this point, the tools that can run code are actually clearer — the script is right there and I can read it line by line. WorkBuddy feels like a black box that hands over the work — fast, sure, but I have to keep my guard up and double-check. For things like submission-and-review forms where I'm accountable once they go out, I never dare send them directly — I always go over them myself first.

My take: for people who deal with documents, spreadsheets, and PDFs every day — work that isn't complex but is high-volume — it's the most cost-effective. Weekly reports, archiving, summarizing, format unification — it really does save time on these, and project-department clerks and records staff fall into this category. But for work that involves professional judgment, it won't do — for quantity takeoffs it can only do rough estimates, and the detailed accounts still have to be done by me; it can't read drawings, and it can't get its hands into the CAD stuff; expecting it to make technical decisions for you is wishful thinking.

The comparison article didn't give a performance ranking in the end either, saying demo results under different environments can't be treated as a ranking — I agree with that. With tools, whether someone else's test shows it's fast is one thing; whether it's handy once it lands in your hands is another.

Don't start by giving it big jobs. First try it on one small thing you do every day and are sick of doing — like renaming a pile of PDFs by date — and once that goes smoothly, add more. I was greedy at first and stuffed in too many requirements at once, and the result was it got confused and so did I.