Header image crafted by AI

Text penned by Tian Yu

Edited by Shen Xiao

On October 14th, WorkBuddy unveiled its deep integration with WeCom.

A new gateway for WorkBuddy has been introduced within WeCom, presenting itself as a WeCom bot in the chat list. This allows users to directly engage with it to streamline their work processes.

Despite my lack of prior experience with WorkBuddy, I embarked on a journey to explore its functions pertinent to daily office tasks from scratch. The outcomes were both astonishing and flawed.

The setup process proved to be straightforward. After installing and logging into WorkBuddy, I located the bot entry in WeCom's chat list, created my personalized bot, and completed the necessary authorization and binding steps to establish a seamless conversation channel.

From that moment onward, whether for routine chats or scheduled tasks, everything seamlessly appeared within this chat window.

On the 15th, WeCom and WorkBuddy co-hosted an official livestream to unveil the highlights and usage techniques of this integration, lasting approximately an hour.

I submitted a recorded video to WorkBuddy, requesting it to summarize the key points of the livestream. While executing the task, WorkBuddy took a considerable amount of time to process, and midway through, it suggested pausing the task and retrying later. Opting to continue would result in repeated consumption of points and run quotas.

During the wait, amusing messages such as "The thinking train has departed, please fasten your seatbelt," "Don't rush, good things come to those who wait," "Almost there, I promise," "My brain is in overdrive," and "I'm racking my brains, and they're all worn out" appeared at the bottom, showcasing Tencent's meticulous attention to detail.

The task could be paused at any point during its execution, taking a total of 14 minutes to complete the summary. Ultimately, the key points output by WorkBuddy were as follows:

Firstly, for beginners, compared to the previous task of "raising virtual lobsters," AI office work now boasts almost zero barriers to entry. Simply upgrade the WeCom desktop client to version 5.0.12, click on the "Smart Bot" entry in the sidebar to claim a trial (limited free quantities, apply to the administrator after depletion).

Secondly, the crux of the integration lies in connecting WeCom's data: messages, files, meetings, emails, schedules, and contacts can all be directly processed by AI.

The official description states that "AI naturally possesses context as soon as it enters the enterprise": if you mention "that project," it knows which project you're referring to; if you mention "the client's meeting," it will retrieve it autonomously. Unlike the previous method of accessing WeCom through a "connector," this time the messaging capability is newly opened, and the product layer has also been deeply intertwined.

However, the open messaging capability still maintains hierarchical levels within the enterprise. Members' access permissions to the context of messages, emails, and documents support tiered control.

In terms of quota configuration, it supports unified purchasing, allocation of slots, and adjustment of quotas through the management backend; models (such as Hunyuan, DeepSeek, etc.) can be uniformly configured in the console, and enterprise-private skills and MCP connectors can be uploaded to integrate the company's internal systems.

During the livestream, the host recommended several usage methods: for instance, asking "What have I been busy with lately," and WorkBuddy directly organizes the main thread of work; multi-device handoff where tasks are assigned on the phone and completed by the time you reach your desk and open your computer; automatic identification of project risks and sending reminders to those responsible; one-click generation of data dashboards from smart tables, etc.

Taking "What have I been busy with lately" as an example, it can clearly list some of my recent work content.

However, since I hadn't left any work traces on WeCom, I was perplexed. Upon further inquiry, it explained in detail: it went through its own conversation records and listed the contents of my desktop and material folders, including a document that was my personal monthly summary.

Just as it said, "The most recent file activity is the most honest answer, possibly even more accurate than your own recollection."

Tencent aspires to elevate WorkBuddy from a mere chat tool to a "digital colleague": users can assign tasks to it, access information and files within WeCom, and then receive the results back in WeCom. The typical scenario showcased by the official is automatically generating a daily report and pushing it to users every day.

I briefly tested an automatic scheduling task: every day at 10:10 AM, automatically collect dynamics in the AI field, filter 3 to 5 valuable pieces of information, and then push them to WeCom.

Problems surfaced during the initial run.

After the task commenced, WorkBuddy indicated that "the current model queue is full." I switched to several models marked as "slightly congested" or "idle" on the interface, but the outcome remained unchanged. From 10:10 AM to 10:47 AM, the task repeatedly attempted and ultimately failed.

"Computing power shortage" has emerged as a barrier to user experience, perhaps a minor hiccup due to early access. The following day, the task was swiftly completed and successfully pushed.

However, users were left puzzled as to why models displaying "idle" on the interface couldn't execute either.

Tencent mentioned in its 2026 corporate governance report that it would continue to bolster its computing power and technological infrastructure; when releasing the official version of Hunyuan Hy3 in July, Caixin reported that Tencent further augmented its post-training computing power scale and focused on fortifying Agent capabilities.

While adjusting the scheduled task, I also requested it to "organize the latest week's AI news and send it to me via WeCom." It promptly replied. However, from WeCom, a temporal hallucination error emerged: the content pushed was marked with a date of October 2025, a full year ahead.

So I inquired about the reason. Initially, it believed it was correct, "The content of this output file is from October 15, 2026."

After I pointed out the issue, it provided a comprehensive explanation: it had experienced a model hallucination. The scheduled task had indeed successfully generated the correct briefing for the 15th, but due to a system bug, it was not pushed to WeCom.

Concerned that this was due to my own operational missteps, I asked WorkBuddy, "Is this my fault in using it, or is it your mistake?" It replied, "It's the AI's mistake."

However, admitting the mistake didn't equate to rectifying it. The next day, I conducted the same test, replicating the previous prompt, and WorkBuddy still pushed a message dated "October 16, 2025."

After reviewing the execution record, I discovered that WorkBuddy had already written the keyword as "October 2025" during the search phase. This wasn't a final-step error in calculating the year; rather, it used the incorrect time from the very outset of the retrieval process.

Only after I explicitly included "October 16, 2026" in the prompt did the result return to normal.

This indicates that there are at least two layers to the time error: one occurs when the model generates content, and the other occurs during task execution and search parameters. Simply reminding the model to "write the accurate date" is insufficient; the system also needs to confirm the current date before retrieval and verify if the content date falls within the user-requested time range before sending.

Moreover, task execution and message pushing are not the same state.

WorkBuddy's official documentation elucidates that local scheduled tasks are executed on the local client, and the computer cannot be closed or the client exited during runtime; WeCom pushing is merely a result notification, and a failed push does not affect the task execution itself nor is it recorded in the assistant's conversation context.

This elucidates the phenomena I observed during testing: WorkBuddy might have generated the file, but WeCom didn't receive it; or the task might have failed to execute, but this wasn't promptly or clearly conveyed to the user.

WorkBuddy rectified issues such as automation tasks displaying "running" for an extended period, normal executions being mistakenly judged as make-up runs, delayed refresh of run records, and idle trigger delays in its September update; the October 16 update mentioned fixing issues where the WeCom bot got stuck in thinking after generating a file, the file wasn't returned, and WeCom notifications for scheduled tasks were delayed or missed.

The direction is correct, but the details necessitate further refinement.

A minor detail was also observed while constantly switching between WeCom and WorkBuddy: messages sent in WeCom are synchronized to WorkBuddy, but conversations in WorkBuddy are not synchronized to WeCom.

Both sides share the same conversation memory, but replies only revert to the origin of the message. This design has its pros and cons: the advantage is that the phone won't be inundated with operations from the computer side; the disadvantage is that if you forget where you posed the question, you might have to search for the answer on both ends.

Computing power is the primary bottleneck, and reliability is the ultimate hurdle. At this stage, utilizing WorkBuddy for time-sensitive automation tasks still necessitates manual verification.

Nowadays, WorkBuddy resembles an intern colleague who has learned to undertake tasks but still requires supervision for delivery. It can orchestrate workflows, but it still falls short of the stability required to truly "let go and trust it."

This integration of WeCom and WorkBuddy has propelled the AI assistant from a "conversational tool" towards a "digital colleague." Colleagues make mistakes too, but as long as the direction is correct, the rest can be left to iteration and refinement.