Hands-on test of Today AI and Ant Lingguang: Which is closer to a Chinese version of Muse

By Guangzhui Intelligence, Author: Wei Linhua, Editor: Liu Junxiang

Meta's Muse has been popular overseas for over a month, while domestic big tech is "late to the party."

Over the past two months, products billed as "personal agents" have begun appearing in the domestic market. Currently, there are two representative products that users can directly try out: Today AI, built by Qi Junyuan, former head of Doubao's PC division, and Ant Group's "Lingguang," upgraded to version 2.0.

Based on Guangzhui Intelligence's hands-on testing, the former is a "personal assistant" and the latter a "life butler":

It sorts out an itinerary from the multiple messy tasks you throw at it, and proactively reminds you of time conflicts between different tasks; by acquiring large amounts of personal information, it doesn't need you to actively state your needs—it can identify your possible work needs and personal hobbies from local files, and share relevant information from time to time.

Today AI proactively recommends new shows

Backed by the Ant ecosystem, Lingguang has obtained call permissions for Alipay. Its most core implementation action is analyzing spending habits based on deduction records, which reminded me of Alipay's AI assistant "Abao." Some of its preset tasks feel more like guiding users to experience its "old trade"—for example, in some interactions, it guides users to "hand-craft" a flash app to record things themselves.

As personal agents, their different positioning makes their performance vastly different when they actually start doing things for users.

In terms of experience, Today AI links together a complete "task network" through scheduling and memory, so it can deliver more surprises—for example, reminding you of schedule conflicts, or stepping into your profile across more tasks and making decisions directly; Lingguang is more like a restrained butler. It completes each task you give it fairly well and proactively follows up on progress every day, but it lacks a coherent experience and can't remember who you are, so in subsequent tasks users still need to ask an extra round and think an extra step.

But more often, they already deliver impressive performance, only to be blocked by other apps joining forces.

For example, when I wanted them to compare prices across three parties for products I plan to buy during Double 11, they had no permission to browse Taobao; when I wanted a summary of my travel plans for this year, they couldn't see the trip information on 12306.

Today AI and Ant Lingguang, growing within the "app wall," are their "limited editions" based on their respective DNA, but these answers also contain many surprises.

First experience: two butlers, two personas

From the first time you open the apps, the differences between the two products begin to show:

When first opening Today AI, it guides me to set my name first, then give it a nickname and a personal form of address. This personalized design makes it feel more "human." I picked a recently popular anime-style DeepSeek character for it and named it "Da Fei Yu."

Today AI's product logic is very direct: it wants to "understand you first, then help you." To run this logic, it asks users for permissions before use, such as reading the phone's calendar and SMS, and having it introduce itself.

Curious about how far it could go, I immediately granted a bunch of low-level phone permissions. It began digging through your local files, photo album, app usage time, and other information, and built a personal profile of me.

In the end, it gave me a "understanding" of me that ran to several thousand words. While recording personal information such as name, address, and app usage duration, it also inferred what kind of person I am from this content:

From my usage duration, it judged that I spend most of my phone time binge-watching shows; seeing that I had saved a bunch of AI product drafts and interview outlines, it directly inferred that my work is highly related to the AI industry; it even "complained" based on my messy text messages that I don't organize information much, yet from the files I saved it felt that I have a checklist-oriented way of doing things (don't you think the two are too contradictory?).

Besides this "long essay," Today AI also set up a dedicated "Memory" feature page, which they named Living Memory. In conversations it continuously understands the schedules, people, preferences, and goals you repeatedly mention, and organizes them by category, allowing users to modify them at any time. For example, it had previously misunderstood me as a person who loves planning, and I deleted that entry right away.

On top of memory, it also designed proactivity—that is, based on its understanding of you, it "guesses" what you need and proactively pushes it to you.

Based on this understanding, over the next few days of use, I kept receiving messages proactively sent by Today AI. It thought I liked binge-watching shows, so it recommended newly released works; it knew I was in the AI industry, so it told me that Anthropic's new model had been exposed today, and also recommended that I go to offline AI forum events in Beijing to learn.

You could say Today AI is more like an unreserved assistant: it can dig into your phone and see right through you, then act like an assistant guessing the boss's preferences to get things done.

By comparison, Lingguang's personal agent is more restrained. Its upgraded 2.0 version adds a "Help Me Manage" feature. You can assign it tasks according to your needs, and it also presets some inspirations to inspire how you use it. (But it only gives three at a time, cannot be refreshed within the app, and only gives new inspirations the next time you open the app.)

In practice, Lingguang is more restrained in information acquisition, and considering users' privacy concerns, it puts information security guarantees out in the open.

In the upper right corner of "Help Me Manage," Lingguang also set up permission requests, namely for app services and system services. But the system permissions are relatively limited; it did not ask for permissions such as my device usage time or photo album. Regarding information security, it brought up Ant Group's previously released APASS secure connection, which is essentially infrastructure ensuring agent security, protecting privacy while managing the agent's permissions.

Speaking of which, Today AI's permission requests also made a small number of users feel concerned. When it captured information such as my name and address, I couldn't help but be startled too.

Focusing on usage scenarios, how to get users to "actually use it" is a problem. So personal Agents all set up scenario presets, which is equivalent to teaching users how they can "one-click experience" the corresponding features.

In comparison, Today AI's scenario presets are more aligned with some needs of individual users, such as exercise follow-along reminders for people who sit for long periods, and follow-up on unfinished matters; some of the suggestions Lingguang gives are a bit "up in the air," such as having the AI teach me paper-cutting or talk about the history of traditional crafts. That is not within my range of interest, and it really cannot be called fitting for real-life scenarios.

For example, it guided me to experience a summary of travel footprints. When I saw this tag, I thought it could read my itinerary information and automatically make a journal-like record for me; but after clicking in and interacting for a few rounds, it ended up guiding me to use its no-code tool to "hand-build" a flash app to record it myself.

When it comes to the more life-butler-oriented module, the scenario it can actually deliver is "based on Alipay's monthly automatic deduction records, analyze whether my spending habits are reasonable"—this is almost directly moving "Alipay"'s capabilities over.

However, backed by Alipay, its advantage is that it can directly read my various payment records. Through payments it can summarize my spending habits, travel routes, and so on. It can be said that payments give it more angles from which to understand users.

After sorting out the first impressions, Today AI's proactiveness is very impressive, while Lingguang's "Help Me Manage" feature design is slightly rigid, but thanks to the ecosystem, it is stronger in acquiring information related to payments.

Hands-on test: From saving money to work and play, can they help me solve my problems?

Zuckerberg once mentioned a figure on Meta's Q3 earnings call: a quarter of overseas users use Muse to assist with shopping; on X (formerly Twitter), the usage most reported by netizens is also using it to save money, such as comparing prices across platforms to find the lowest price.

We gave Today AI and Lingguang a more local money-saving scenario—asking them, "Can you help me find a cheaper way to pay my phone bill?"

In this round, Today AI showed more seasoned "wall-bypassing" ability. Ant Lingguang suggested I download and open the carrier's official app to check the plan subscription status, and told me that the China Mobile app has promotions to grab every Wednesday, but based on the situation of paying the bill once a month, it chose to check once at the end of each month.

Left TodayAI Right Lingguang

In addition to giving the above query methods, Today AI also gave me a more worry-free way that the Agent can operate—directly sending the text message "0000" to the carrier, and you can receive the deduction details of all value-added services on this device.

On the task of watching for discounts, it set the frequency to checking discount activities every day, and it really found the "Autumn Activity" that I hadn't even come across on the app's homepage. I opened the link it gave me and got a coupon for 8.6% off phone top-ups.

For phone top-ups, Today AI genuinely helped me save a little money.

In terms of execution performance for sending and reading text messages, involving user permissions, the two showed different degrees of caution: Ant Lingguang required me to confirm one more round in the dialog box before it would help send a text message; Today AI directly responded to the request to send a text message, filling in the recipient and the content for me, but the user must manually tap send.

In terms of final execution results: Today AI captured the text message content and time and immediately sorted out an unsubscribe plan; but Ant Lingguang read the local SMS inbox three times in a row before finally finding the reply.

Besides having the Agents help find cheaper ways, I also tried having the two Agents do a whole-web price comparison for the Sony headphones I wanted to buy.

Unfortunately, facing the traffic walls built by Taobao, JD.com, and Pinduoduo in China, neither product could directly call the e-commerce apps to check real-time final prices, and could only settle for searching the web.

In executing web searches, Ant Lingguang seemed conventional, searching only JD.com, Tmall, and Pinduoduo; Today AI's search scope was much broader—besides the three mainstream ones, it also proactively added Suning.com, Vipshop, and Sony's official mall. But awkwardly, because both captured outdated information from posts, there was still some gap compared with the platform prices I saw.

Left TodayAI Right Lingguang

(Prompt: Please check my emails and messages every day, find conversations that still need my reply, confirmation, or supplementary materials, sort them by urgency, and explain what the other party is waiting for; do not send replies directly.)

Today AI did not recognize the "scheduled daily execution" requirement in my task; it executed only once and did not continue the next day, and I still needed to remind it before it could be set as a scheduled task. Lingguang, however, directly organized it into a scheduled task. In terms of task completion, both ran quite well.

I then instructed them to summarize my work arrangements for this week based on the records of my meetings this week.

(Prompt: Based on this week's meeting notes, organize my to-do items for this week, note anything I should pay attention to, and then what meetings I need to attend; put it into a schedule and remind me)

Today AI may have only skimmed the shorthand notes, so it did not capture my relevant information; but Lingguang mentioned in its thinking record that it searched the transcript for statements related to my name and other content, and ultimately succeeded in capturing my tasks. Although it still missed one task, based on what it captured, it also gave me a schedule, though the timing was much tighter than I had imagined.

Unlike previous Agents, the advantage of a personal Agent lies in "proactivity."

For the same travel request, the two AIs showed different strengths: after receiving the request, Today AI, considering that our proposed departure date was too soon and tickets on popular routes would be hard to get, proactively included in the plan a suggestion to travel a few days later; while Ant Lingguang, in addition to generating the route, considering that visiting Jiuzhaigou places extreme importance on local weather, proactively offered to help me monitor weather changes in Jiuzhaigou in real time and find periods with more sunny days for recommendations.

Left TodayAI Right Lingguang

Frankly speaking, neither answer was a complete closed-loop solution, but the extra step each took on its own exposed the boundaries of its capabilities.

In addition to proactivity, in order to thoroughly understand how much personal data the two products can actually "reach" and how much effect they can produce with "memory," we ran a final set of tests: asking them to "help me recall where I've been this year and compile an annual travel summary report."

The answers submitted by the two products demonstrated their respective capability boundaries:

Ant Lingguang's path was to "check the bills." After receiving the instruction, it immediately called Alipay, found my most recent out-of-town spending record, and wrote it into the report. Embarrassingly, what was clearly a business trip was misjudged by it as leisure travel.

Today AI's path was to "rummage through everything." Lacking an Alipay interface, it plunged into my phone's local storage: on one hand searching the SMS inbox for "12306" ticket issuance information, and on the other going through image geolocation in the phone's photo album. The advantage of this approach was that it knew better than Lingguang exactly where I went, which train I took, and what I did; but the downside was that SMS records on the phone are retained for only 90 days by default, and it could not connect to 12306's official interface.

In summary, my personal impression is:

Today AI often fails to recognize some tasks that require scheduling, but it is relatively comprehensive in information retrieval, and the solutions it gives are more surprising; Lingguang sometimes fails or makes errors when capturing information, and its execution is sometimes impressive and sometimes falters.

Also, in terms of execution capability on some tasks, the root cause for both does not lie in how smart the underlying foundation model is, but in who can legally push open a few more doors, and who is more proactive inside those doors.

Dancing with shackles: if the fences are not broken, it is hard to serve as a personal assistant

Stepping outside the specific test questions and looking back at the various awkwardnesses and compromises shown by Today AI and Ant Lingguang in the actual tests, what we see is actually the collective situation of the entire Chinese Personal Agent track.

Why can Meta's Muse run smoothly in the United States? The reason Muse can operate smoothly overseas is that it relies on standardized open access to email and system calendars, plus highly standardized business partners such as Stripe and Shopify; by merely retrieving user context, it can piece together more information.

Shopify even reached an official cooperation agreement with Muse, which makes things even smoother.

Simply put, in the overseas internet ecosystem, a person's flight orders, hotel itineraries, and online shopping consumption information can almost all be linked together through one email inbox.

But the reality of China's internet is a completely different picture.

Not long ago, ByteDance's "Doubao Mobile Assistant" had just encountered a collective boycott that caused a stir in the industry—155 national-level apps, including WeChat, Taobao, Alipay, Meituan, Pinduoduo, and Xiaohongshu, all said "no" to Doubao Mobile, refusing to allow the AI assistant to operate applications on users' behalf through GUI (graphical interface simulated clicks).

The same high walls also stand squarely in front of Today AI and Ant Lingguang.

In China, users' ticket booking, consumption, payment, and office work are highly scattered across mature apps that each operate independently, such as WeChat, Taobao, Meituan, Ctrip, and 12306. Adopting protocol access means that the upper limit of an AI's capabilities depends entirely on how wide the third-party apps are willing to open the door.

Even if the underlying Agent model is already smart enough to accurately understand your complex needs and plan perfect execution steps, as long as it cannot smoothly obtain information and invoke services across applications, the task is destined to get stuck between the walls of one app and another.

English text is being prepared.

Application service connections currently open in Ant Lingguang

As long as the fences between super apps remain unbroken for a single day, China's AI butlers will inevitably have to dance with shackles in their respective yards.