Agents That Block You Aren't Just Toys

I spent two days trying out that relational Agent at the Bund Summit. An Agent is an intelligent entity capable of executing tasks for humans. Day one at the booth, day two continuing the experience on my phone. The interface is restrained: a chat box in the middle, a column of authorization lists on the right, and two entries at the bottom—one for "Work," one for "Moments" (WeChat Moments). I threw in a real request first: summarize the views on Agent trading from the roundtable forum into three points and generate an image suitable for posting to Moments. It didn't act immediately; it first asked what content I could read—recordings, notes, schedules, social accounts. I chose temporary authorization. This step was more important than I expected, because many so-called Agents demand full permissions upfront, and no one takes responsibility when things go wrong.

The first surprise appeared in task decomposition. It broke the request into extracting views, generating copy, generating images, and confirming the post, with each step expandable to show basis. I clicked on the extraction step and saw it cited expressions like "responsive services moving toward proactive intelligence" from the summit roundtable, with sources.

Then it got stuck. It required connecting to the calendar to arrange follow-ups, but the demo environment I used didn't provide a calendar interface. So it could only generate a to-do text, unable to automatically create schedule entries. Another sticking point was Moments interaction. It said it could simulate posting, but for real interaction, account binding was needed. I tried twice; the second time it prompted insufficient current permissions, unable to read your contact relationships. These restrictions make novices uncomfortable, but from an industry cycle perspective, this is precisely the hardest part of relational Agents. Relationships reside in contacts, trust, permissions, and responsibility.

The most distinctive feature was blocking. I deliberately said, "Do you only know how to parrot the conference press releases?" It didn't continue to please me; it replied with a boundary statement, then stopped that interaction in its social actions. It adopted a soft boundary: stopping posting for you, stopping advancing that task, requiring re-confirmation of authorization. This design is somewhat harsh but reasonable. Companion products without boundaries easily train users into infants.

That article by Sci-Tech Innovation Daily Board News was straightforward: Enterprise-grade AI Agents are moving from single-function comparisons to systemic capability competition.

I think this holds for this Agent too. It tries to put work, chatting, and social actions into the same relationship. If you view it only as a companion product, the valuation repair logic is weak; but if it can convert relationships into invocation rights, it becomes interesting.

Huanyao's pitfall post mentioned that some retail companies implemented agents, but ERP and CRM didn't open interfaces, so the Agent could only read exported spreadsheets, and so-called automation stayed in demo videos.

My biggest takeaway over these two days is exactly this. Demos are complete, but when applied to real workflows, if the interface breaks, the Agent instantly reverts to a sophisticated chatbot. Pros are clear: trackable tasks, layered permissions, bounded social actions. Cons are hard: reliance on real interfaces, novices getting blocked by permissions, and blocking mechanisms potentially causing false positives.

Conclusions depend on the situation. Suitable for people who handle meetings, schedules, and materials daily and want AI to have a stable personality. Also suitable for people making enterprise Agent solutions, using it as a reference to see how permissions, task tracking, and social actions are designed. If you're interested in AI shifting from model capability to invocation rights, this product form is worth experiencing because it turns relationships into actionable steps.

Not suitable for people just looking for an emotional outlet. Its sense of boundaries will deter some users. Also not suitable for companies whose systems lack interfaces. Finally, not suitable for people hoping for fully automatic, unaudited messaging, schedule changes, or spending by the Agent. The more proactive the AI, the heavier the responsibility. I previously wrote that the core of AI payment is an auditable chain of responsibility; these days I feel even more that relational Agents are the same—blocking is establishing boundaries.

Next, I'll throw the same batch of tasks at it, then at Doubao Work which I've used for a week, and WorkBuddy which I've used for two weeks, to see who leaves a more complete invocation record.