For the reason that basic AI agent Manus was launched final week, it has unfold on-line like wildfire. And never simply in China, the place it was developed by the Wuhan-based startup Butterfly Impact. It’s made its method into the worldwide dialog, with influential voices in tech, together with Twitter cofounder Jack Dorsey and Hugging Face product lead Victor Mustar, praising its efficiency. Some have even dubbed it “the second DeepSeek,” evaluating it to the sooner AI mannequin that took the trade unexpectedly for its sudden capabilities in addition to its origin.
Manus claims to be the world’s first basic AI agent, leveraging a number of AI fashions (comparable to Anthropic’s Claude 3.5 Sonnet and fine-tuned variations of Alibaba’s open-source Qwen) and varied independently working brokers to behave autonomously on a variety of duties. (This makes it completely different from AI chatbots, together with DeepSeek, that are based mostly on a single giant language mannequin household and are primarily designed for conversational interactions.)
Regardless of all of the hype, only a few individuals have had an opportunity to make use of it. At the moment, below 1% of the customers on the wait record have acquired an invitation code. (It’s unclear how many individuals are on this record, however for a way of how a lot curiosity there may be, Manus’s Discord channel has greater than 186,000 members.)
MIT Expertise Overview was capable of get hold of entry to Manus, and once I gave it a test-drive, I discovered that utilizing it appears like collaborating with a extremely smart and environment friendly intern: Whereas it sometimes lacks understanding of what it’s being requested to do, makes incorrect assumptions, or cuts corners to expedite duties, it explains its reasoning clearly, is remarkably adaptable, and might enhance considerably when supplied with detailed directions or suggestions. In the end, it’s promising however not excellent.
Similar to its guardian firm’s earlier product, an AI assistant referred to as Monica that was launched in 2023, Manus is meant for a world viewers. English is about because the default language, and its design is clear and minimalist.
To get in, a consumer has to enter a sound invite code. Then the system directs customers to a touchdown web page that carefully resembles these of ChatGPT or DeepSeek, with earlier periods displayed in a left-hand column and a chat enter field within the heart. The touchdown web page additionally options pattern duties curated by the corporate—starting from enterprise technique growth to interactive studying to personalised audio meditation periods.
Like different reasoning-based agentic AI instruments, comparable to ChatGPT DeepResearch, Manus is able to breaking duties down into steps and autonomously navigating the online to get the knowledge it wants to finish them. What units it aside is the “Manus’s Laptop” window, which permits customers not solely to watch what the agent is doing but in addition to intervene at any level.
To place it to the check, I gave Manus three assignments: (1) compile an inventory of notable reporters masking China tech, (2) seek for two-bedroom property listings in New York Metropolis, and (3) nominate potential candidates for Innovators Below 35, an inventory created by MIT Expertise Overview yearly.
Right here’s the way it did:
Process 1: The primary record of reporters that Manus gave me contained solely 5 names, with 5 “honorable mentions” beneath them. I seen that it listed some journalists’ notable work however didn’t do that for others. I requested Manus why. The rationale it provided was hilariously easy: It acquired lazy. It was “partly attributable to time constraints as I attempted to expedite the analysis course of,” the agent informed me. After I insisted on consistency and thoroughness, Manus responded with a complete record of 30 journalists, noting their present outlet and itemizing notable work. (I used to be glad to see I made the lower, together with a lot of my beloved friends.)
I used to be impressed that I used to be capable of make top-level options for adjustments, a lot as somebody would with a real-life intern or assistant, and that it responded appropriately. And whereas it initially ignored adjustments in some journalists’ employer standing, once I requested it to revisit some outcomes, it shortly corrected them. One other good function: The output was downloadable as a Phrase or Excel file, making it simple to edit or share with others.
Manus hit a snag, although, when accessing journalists’ information articles behind paywalls; it steadily encountered captcha blocks. Since I used to be capable of observe alongside step-by-step, I may simply take over to finish these, although many media websites nonetheless blocked the instrument, citing suspicious exercise. I see potential for main enhancements right here—and it could be helpful if a future model of Manus may proactively ask for assist when it encounters these kinds of restrictions.
Process 2: For the condominium search, I gave Manus a posh set of standards, together with a finances and a number of other parameters: a spacious kitchen, outside area, entry to downtown Manhattan, and a significant prepare station inside a seven-minute stroll. Manus initially interpreted obscure necessities like “some type of outside area” too actually, fully excluding properties with no personal terrace or balcony entry. Nevertheless, after extra steering and clarification, it was capable of compile a broader and extra useful record, giving suggestions in tiers and neat bullet factors.
The ultimate output felt straight from Wirecutter, containing subtitles like “greatest general,” “greatest worth,” and “luxurious choice.” This activity (together with the back-and-forth) took lower than half an hour—rather a lot much less time than compiling the record of journalists (which took a bit over an hour), possible as a result of property listings are extra brazenly obtainable and well-structured on-line.
Process 3: This was the biggest in scope: I requested Manus to appoint 50 individuals for this 12 months’s Innovators Below 35 record. Producing this record is a gigantic endeavor, and we usually get a whole lot of nominations yearly. So I used to be curious to see how nicely Manus may do. It broke the duty into steps, together with reviewing previous lists to grasp choice standards, making a search technique for figuring out candidates, compiling names, and making certain a various collection of candidates from all around the world.
Growing a search technique was essentially the most time-consuming half for Manus. Whereas it didn’t explicitly define its strategy, the Manus’s Laptop window revealed the agent quickly scrolling by web sites of prestigious analysis universities, bulletins of tech awards, and information articles. Nevertheless, it once more encountered obstacles when making an attempt to entry tutorial papers and paywalled media content material.
After three hours of scouring the web—throughout which Manus (understandably) requested me a number of occasions whether or not I may slender the search—it was solely capable of give me three candidates with full background profiles. After I pressed it once more to offer a whole record of fifty names, it will definitely generated one, however sure tutorial establishments and fields had been closely overrepresented, reflecting an incomplete analysis course of. After I identified the problem and requested it to seek out 5 candidates from China, it managed to compile a stable five-name record, although the outcomes skewed towards Chinese language media darlings. In the end, I had to surrender after the system warned that Manus’s efficiency may decline if I saved inputting an excessive amount of textual content.
My evaluation: General, I discovered Manus to be a extremely intuitive instrument appropriate for customers with or with out coding backgrounds. On two of the three duties, it offered higher outcomes than ChatGPT DeepResearch, although it took considerably longer to finish them. Manus appears greatest suited to analytical duties that require in depth analysis on the open web however have a restricted scope. In different phrases, it’s greatest to stay to the types of issues a talented human intern may do throughout a day of labor.
Nonetheless, it’s not all clean crusing. Manus can endure from frequent crashes and system instability, and it could battle when requested to course of giant chunks of textual content. The message “Because of the present excessive service load, duties can’t be created. Please strive once more in a couple of minutes” flashed on my display just a few occasions once I tried to start out new requests, and sometimes Manus’s Laptop froze on a sure web page for a protracted time period.
It has a better failure price than ChatGPT DeepResearch—an issue the group is addressing, in accordance to Manus’s chief scientist, Peak Ji. That mentioned, the Chinese language media outlet 36Kr studies that Manus’s per-task price is about $2, which is simply one-tenth of DeepResearch’s price. If the Manus group strengthens its server infrastructure, I can see the instrument changing into a most popular alternative for particular person customers, notably white-collar professionals, impartial builders, and small groups.
Lastly, I believe it’s actually precious that Manus’s working course of feels comparatively clear and collaborative. It actively asks questions alongside the way in which and retains key directions as “information” in its reminiscence for future use, permitting for an simply customizable agentic expertise. It’s additionally very nice that every session is replayable and shareable.
I count on I’ll hold utilizing Manus for all kinds of duties, in each my private {and professional} lives. Whereas I’m undecided the comparisons to DeepSeek are fairly proper, it serves as additional proof that Chinese language AI corporations aren’t simply following within the footsteps of their Western counterparts. Fairly than simply innovating on base fashions, they’re actively shaping the adoption of autonomous AI brokers in their very own method.
