
Every AI agent keeps two histories. The first is the one you can see: the chat transcript, the task list, the finished document sitting in your downloads folder. The second is quieter, and it is the one that caught a group of Chinese developers by surprise this week. When users of Tencent's office agent WorkBuddy went looking in the app's own log folder, they found records that went well beyond a tidy list of errors — discussion on LINUX DO centered on the discovery that the tool records what users send it, not just whether it crashed.
That surprise is worth unpacking, because it is about a category shift: an office agent is not a chatbot in a browser tab. It lives on your machine, reaches into files and messaging apps, runs tasks while you are away, and keeps diagnostic traces of all of that. The question is not whether it logs things, but which layers exist, who can read them, and what you should stop pasting into the box.
WorkBuddy is Tencent Cloud's AI-powered desktop workstation for office work: you describe a goal in plain language, and it plans the steps, calls tools, and hands back a finished deliverable such as a report or slide deck. According to Tencent's own announcement on 29 October 2026, it launched in March 2026 and has since handled more than nine billion task requests, across more than 50 industries and half of China's top 500 enterprises, with an ecosystem of more than 50,000 developers and partners and more than 200,000 skills.
That scale matters for the logging question: a tool inside finance, marketing, and human-resources departments cannot afford vague data practices. The good news is that WorkBuddy's own documentation is unusually explicit about where its logs live — which makes it a useful case study for auditing any agent, whether or not you ever install this one.
Most users think of "history" as a single thing. In practice, an agentic desktop tool tends to keep three separate layers, and each one answers a different question:| Layer | What it typically holds | Who usually sees it | |---|---|---| | Chat and task history | Your prompts, the agent's replies, files it produced | You, in the app | | Local diagnostic logs | Timestamps, tool calls, errors, and — in some cases — message content and device details | Anyone with access to your machine, plus support if you upload them | | Server-side traces | The request as processed by the model provider, usage counts, and feedback submissions | The vendor, under its privacy policy |
The LINUX DO discussion, summarized here rather than quoted, focused on the middle layer: the local log files. That is the layer people forget, because nothing in the chat window advertises it. It exists for a legitimate reason — when an agent fails halfway through a multi-step task, support needs more than "it didn't work" — but its contents can be richer than the word "log" suggests.
WorkBuddy's own FAQ confirms the shape of this. It tells users where to find the logs — on macOS, Help then Open Log Folder; on Windows, the Help menu then Open Log Directory — and explains that feedback involves an explicit Upload Logs checkbox. The FAQ then adds the sentence that should anchor any discussion of this product: logs are used solely for troubleshooting and may contain data such as conversation records and device information. That is not a leak or a secret feature. It is the vendor's own description, and exactly the kind of line worth reading before you hand an agent sensitive work.
If the documentation says it, why did the forum thread take off? Almost nobody reads a troubleshooting FAQ before installing a productivity tool, and most people carry a chatbot mental model: you type, you get an answer, and the record is the transcript you can see and delete. An agent breaks that model in three ways.
First, it acts. A chatbot answers; an agent opens files, queries internal systems, messages colleagues through connected apps, and chains those steps together. Each action generates a trace that is more revealing than the final answer, because it shows the intermediate steps and the data the agent fetched.
Second, it persists. WorkBuddy's FAQ notes that its Assistant can receive and execute tasks remotely via Slack, Telegram, and Discord — but only while the app is running on your computer. That is a real boundary, and it also means the machine holding the logs is the machine doing the work. Anyone who later opens that folder — a colleague borrowing the laptop, an IT drive image, a forgotten support upload — sees more than your chat history.
Third, it blurs work and personal use. The same agent that drafts a budget model in the morning can summarize a personal document in the afternoon. If both sessions land in the same logs and the same feedback upload, the separation you assumed never existed at the log level. Forum users were, in effect, noticing that the receipt drawer is shared even when the conversations feel separate.
None of this makes WorkBuddy unusual. If anything, its documented log location and feedback checkbox put it ahead of tools that bury the same behavior. The surprise is a property of the category, and the fix is a habit, not a boycott.
You can run this audit on WorkBuddy today, and the same steps transfer to almost any desktop agent.
1. Find the log folder before you need it. Open the app's Help menu and locate the log directory, as WorkBuddy's FAQ describes. Check file names, sizes, and dates: a folder that grows every session tells you something about volume before you read a line.
2. Read the feedback path, not just the privacy policy. Policies describe categories of data; feedback flows describe transfers. In WorkBuddy, feedback goes through Settings, then Help and Feedback, and uploads logs only if you check the box. Check whether your agent has an equivalent toggle, whether it is on by default, and whether uploads include attachments you did not intend to send.
3. Separate the machines or the accounts, not just the chats. If you use an agent for both work and personal tasks, assume the logs mix. Options are a separate user account, a separate machine for sensitive work, or a rule that genuinely sensitive material never goes through the agent. Deleting a chat thread is not the same as clearing diagnostic logs, and neither is the same as deleting a server-side record.
4. Map the remote-control surface. Any agent that accepts tasks from a messaging app has an inbound path worth understanding: which platforms are connected and who can message the bot. WorkBuddy's answer is clear — Slack, Telegram, and Discord, only while the app is running — but the general question applies everywhere: an agent you can reach from your phone is one someone else might reach if permissions are loose.
5. Check the model choice, if there is one. Tencent says WorkBuddy customers can choose the model behind each task. Different providers have different retention and training policies, so treat model selection as a data-routing decision, not just a quality knob.
The audit tells you where data goes. This list tells you what should not go there in the first place:
There is a simple test worth adopting before you give any new agent real work. Ask its documentation three questions: where do the logs live on my machine, what exactly can they contain, and what happens when I press the feedback button? If you cannot answer all three in under ten minutes, you have not adopted a tool yet; you have adopted a mystery.
WorkBuddy passes the first part of that test unusually well: its FAQ names the menu paths, names the contents in plain language, and names the checkbox that controls uploads. The LINUX DO discussion is best read as users holding the product to its own documentation — and discovering that "logs" is a bigger word than they assumed. Every agent keeps receipts. The users who stay in control are the ones who have opened the drawer, looked inside once, and decided what belongs there before the next task runs.
For a related angle on the same vendor's developer tooling, see how Chinese developers wire third-party models into Tencent's CodeBuddy CLI.