Honor has a new AI agent who can read and understand your screen

He chose a restaurant, but could not complete the process because the chosen place required a credit card to confirm the reservation, at which time the user had to take over. You can be flexible in your consultation: in another example, ask him to reserve in a “highly valued” restaurant meant that he would look for opinions with high scores, although the agent does not investigate more than that. It does not cross the opinions of opening with data from other parts of the web, especially since all these data are processed on the device and are not sent to the cloud.

This type of agricultural artificial intelligence is the word of fashion in the technological field. My colleague Will Knight recently tested an assistant of AI who could navigate online and perform online tasks. At the end of last year, Google presented its GEMINI 2 AI model, trained to perform actions in your name. It also renews the idea of ​​a generative user interface for smartphones: In the MWC 2024, we saw some companies working on ways of interacting with applications without using them at all, instead of relying on AI attendees to generate a user interface as you issued a command.

The honor approach resembles in some way what Rabbit (of the infamous Rabbit R1) is doing with Teach mode, in which you train its assistant manually to complete a task. There is no need to access the application programming interface (API) of an application, which is the traditional way in which applications or services communicate with each other. The agent memorizes the process, which allows him to issue the order and make him execute the task.

But Honor affirms that his self -sufficient execution model is not trained to follow strict steps: he is able to recognize the multimodal context of the screen to perform tasks autonomously. Instead of having to train the assistant to learn each and every one of the parts of the opening application, he is able to understand the semantic elements of the user interface and follow a several steps process to execute his application. Honor stressed that this process was more profitable: “Unlike competitors such as Apple, Samsung and Google, which depend on external API (which translates into higher operational costs), the honor agent independently manages a wide range of tasks.”

Other Honor Launches

Although Honor states that its user interface agent uses its own execution models, it also takes advantage of the GEMINI 2 Google Language Language Model, which is what drives the recognition of the intention of its command and the “improved semantic understanding” of what appears on the screen. Google has not given details about the nature of collaboration.

Honor states that it has also been associated with Qualcomm to conserve the data on the device and develop a personal knowledge base that learns your preferences over time. The idea is that if you tend to ask for certain types of food in a delivery application, if you ask the agent to ask in your name, use that context to choose something you know you like. The company states that it is already using some of these AI agents in China.

In his KeynoteHonor also announced that he will offer seven years of updates of software For its Magic 7 Pro flagship and the next devices, matching the Google and Samsung software update policies for Pixel and Galaxy phones.

He also presented a handful of new Gadgets At the fair, including headphones Honor Earbuds Openthe smart watch Honor Watch 5 Ultrathe Tablet Honor Pad V9 and the laptop Honor Magicbook Pro 14. These devices will not be sold in the US, such as most honor products, but will be available in other markets. (The brand received Wired in its media event at the MWC 2025 and paid part of the travel expenses of our reporter).

Article originally published in Wired. Adapted by Mauricio Serfatty Godoy.