Agents Have Eyes Now: A Hands-On Test of Teamily AI's Video Agent
Teamily AI has launched a video-call Agent. The Agent now has a “face” and can also see the user’s world through the camera and AI glasses. Hands-on test: speak your requirements to the camera, get a list of 52 projects in 4 minutes, expanding to 237 after 20 minutes. Feishu, DingTalk, and the rest are going to lose sleep over this.
Teamily AI has launched a video-call Agent. The official description is blunt: every Agent now has a “face,” and can see the world you see through your camera and AI glasses.
What is this experience like? Let’s look at a real case.
A netizen, @叫我阿杭, recently has been leading a team doing affiliate marketing and has been looking for more good affiliate programs. He directly called the Agent via video, held his laptop up to the camera, and talked while letting the Agent watch. The Agent turned the call content into meeting notes and got to work based on the notes.
Official launch video:
In 4 minutes, the first version of the affiliate marketing information database was delivered: 52 projects, with commission rates, payout cycles, cookie duration, payment methods, and suitability for domestic creators itemized.
He asked the Agent to add more information sources and make it automatically organize every day. 20 minutes later, the second version was delivered: 4 information sources, 237 projects, 185 newly added, 68 curated recommendations, and even how to set up the automated loop was written into the instructions.
User test video:
Video calls are only the surface. The real change is that the Agent has gone from “taking instructions” to “being able to see.” For scenarios that are hard to describe in words, just show it directly.
Reactions in the comments vary. @叫我阿杭 thinks the product and research teams at Feishu, DingTalk, and WorkBuddy won’t be able to sleep tonight, and some netizens say this is a paradigm shift; some mention that the Doubao team is also paying attention, that voice and video capabilities are already available and mature, and that they should catch up soon; others are not convinced, thinking it only wastes tokens and bandwidth and can’t see what use it has.
Whatever the judgment, one trend is clear: Agents are beginning to have visual capabilities, and not limited to video calls.
@OmAI_lab’s open-source vision model VLX-Seek takes another direction. In visual recognition, previously every new requirement might require re-labeling data and training a model. VLX-Seek tries to change this: with the same image, change the sentence and you can specify different localization targets.
“Find the black package surrounded by white courier bags”
“Find the worker on the left wearing blue work clothes”
“Find the blue bucket under the worker in the white hat”
Demo video:
Previously, people had to adapt to the model’s detection categories; now, the model is made to understand people’s needs. Industrial quality inspection, warehousing and logistics, robotics, smart hardware—many visual recognition scenarios that were hard to realize in the past now have new approaches. The model code and weights are open-sourced on GitHub.
Some netizens in manufacturing said in related discussions that they had researched this kind of technology before but were too busy; visual monitoring can check equipment progress and feed it back to computers in time, which is a real need.
Agents can “see” now. Next, it’s time to see how Feishu, DingTalk, and the rest respond.
发布时间: 2026-11-20 18:13