Tencent Tech "AI Future Compass" Special Contributor ｜ Xiaojing
Editor ｜ Mengmeng
In the first half of 2025, Agents have become one of the most discussed topics in the large model field.
In this wave of Agents, products have already formed two major camps: vertical Agents focusing on specific domains, and general Agents attempting to cover all scenarios. This debate about "who is the ultimate form of Agent" might be premature—the underlying model capability is the real bottleneck for product capability: general-purpose agents are far from omnipotent, and the depth of vertical agents is also limited.
For user decision-making currently, what's more crucial is: Can Agents accurately integrate into workflows? Is the value provided worth the user's investment? After the trial period, will users spontaneously recognize their indispensability?
Putting aside the technical route debate, let's return to a practical perspective. Tencent Tech will conduct hands-on tests of the three most popular products currently: Manus, Flowith (Agent Neo), and Lovart, demonstrating their best use cases to provide readers with firsthand practical references.
Quick Summary
Understanding the Differences Between the Three Products
First, these three Agent products have distinct positioning differences:
Although Manus and Flowith are both general-purpose Agents, Manus is more like a "digital colleague" that can independently deliver complete products. Its main feature is to directly channel any idea through a complete toolchain of browsers, terminals, and code editors, automatically breaking it down into subtasks and running them to completion.
Flowith (Agent Neo), on the other hand, emphasizes "visual collaboration" and infinite steps: a single conversation in an infinite canvas can spawn multiple parallel threads, allowing team members to drag and drop materials, comments, and branches freely. Oracle Mode dynamically reorders priorities and completes the creation of websites, mini-programs, or 3D interactive pages over thousands of steps.
Lovart is deeply specialized in design scenarios. It acts like an outsourced studio, breaking down user requirements into four pipelines—theme, style, materials, and layout—then calling multi-modal models for image, animation, and sound to produce logos, posters, short videos, and even print die-cuts all at once, while maintaining editable layers for direct import into Figma and PS for further refinement.
Secondly, in typical application scenarios:
Manus's "delivery-oriented" capability is most suitable for knowledge work requiring complete deliverables: market research reports, financial models, and lengthy legal memorandums can all be completed in a matter of minutes, complete with citations and source files.
Flowith excels in creative scenarios with large amounts of information and requiring multi-person iteration: for example, we import thousands of documents or social media data into a "knowledge garden." The system retrieves and annotates in real-time, presenting node relationships on the canvas. Developers can also collaborate directly on the same plane, creating a closed loop from "idea-sketch-product" within one interface.
Lovart's focus is on high-value brand visual and content marketing. For instance, a creator can request, with a single sentence, "a launch promotion plan for a plant-based skincare brand on a certain social media site." It then outputs five posters and a 30-second teaser video, all with brand colors, layout templates, and social media dimensions, at the click of a button.
We have summarized the key information in the table below:
Application Test
Scenario 1: Simple Creative Scenarios, GPT-4o Might Be Better
In this scenario, we used two prompts to compare the display effects in simple creative scenarios.
Since Lovart and Manus both use GPT-4o as their underlying model, for the completeness of the test, we also used GPT-4o for comparison.
From the final generation results, they are equally matched, and the styles are very similar. However, in terms of image quality and mixed text-image layout, Lovart outperforms Manus, Flowith, and GPT-4o. But GPT-4o's generation speed is faster than all three Agents.
Prompt 1:
Concise and creative advertisement, with a clean white background.
A real [robot] is integrated into hand-drawn black ink graffiti, with smooth lines and full of fun. The graffiti depicts [AI Real]. Add bold black "[We Are Friends]" text at the top. Clearly place the [AI] logo at the bottom. The visual effect should be concise, interesting, high-contrast, and conceptually clever.
Below are the test results:
Prompt 2:
Design an image in the style of colossus worship, aspect ratio 9:16, mysterious style, and with a sense of story.
Below are the test results:
Scenario 2: Drawing Complete Comic Strips Based on Storyboards, Manus Has More Consistent Style
The three Agents generate images with similar styles because they use the same underlying models.
Among them, Lovart has the strongest Chinese text-image mixing capability, with clear text and harmonious display within the overall composition; Manus's text sometimes overlays the image, with some Chinese characters appearing scrambled; Flowith's overall performance is decent; the base model GPT-4o shows scrambled Chinese characters.
In terms of style consistency, Manus and GPT-4o are the most consistent, followed by Flowith.
Scenario 3: Using English Prompt to Output Comprehensive Creative Scenarios: Flowith Goes All Out
This time, we tested with an English Prompt, asking the Agent to output a complete digital baroque style design based on the prompt's requirements, with application scenarios for brand promotion, online virtual clothing promotion, and social media promotion. Below is the original Prompt, which specifies content theme, style, colors, texture, and size.
Prompt:
Content:
Theme: "Digital Luxury and Future Fashion." Combines Digital Baroque and trendy fashion to showcase virtual clothing, digital models, and futuristic accessories, conveying a bold and luxurious brand identity.
Style:
A mix of Digital Baroque and modern trends with intricate 3D patterns, futuristic geometric lines, and dynamic lighting.
Color Palette:
Deep metallics (rose gold, bronze) + vibrant neon hues (electric blue, fluorescent purple, neon pink) for strong contrast.
Mood/Material:
Dreamy, surreal, with materials like digital metallic textures, holographic gradients, and glossy glass effects.
Proportion/Size:
16:9 HD horizontal, optimized for mobile screens.
Usage Scenario:
Perfect for digital fashion branding, virtual clothing promotion, and social media campaigns.
Lovart returned three images: a model wearing baroque-style clothing, baroque-style accessories, and a baroque-style background. In the final interface, we can also choose to re-edit the images. The toolbar at the top allows for actions such as image expansion, background removal, and adding certain accessories.
Figure: Lovart's output
Manus's final deliverable was only one image and text description, but the image quality was good, and the style matched digital baroque.
Figure: Manus's output
In this test, Flowith's output exceeded expectations. It went all out, providing a super detailed analysis of the digital baroque style in a free canvas, ultimately generating 10 images, an interactive website page, and the website's source code.
Figure: Flowith's output
Figure: Website generated by Flowith
Scenario 4: Comprehensive Scenarios: Both Have Strengths and Weaknesses
Given a comprehensive task to the Agent, guide it to complete complex objectives with a single prompt. Task example: Simulate a startup beverage company, providing only one brand positioning slogan, avoiding excessive information interference, and requiring the Agent to complete all content listed in the prompt within a single conversation.
Prompt:
Role: You are the Brand Design Director.
Goal: Using the slogan "Galaxy Zest – Bottling the Starlit Fizz" to deliver the following:
1. Generate a Logo, a three-color palette, and two A3 vertical poster layouts;
2. Create a 15-second vertical intro video, including brand slogan animation and original music;
3. Output a TikTok release script (3 segments) and a 7-day release schedule;
4. Provide Figma/PSD source file download links.
Manus output a compressed file package, which included the brand logo and color scheme, A3 vertical posters, a 15-second intro video, a dynamically displayed webpage (HTML5 version), and a TikTok marketing plan.
However, it only provided Figma/PSD structure descriptions, without downloadable source files or links. This requirement is actually quite complex; Figma and Photoshop source files are like complex "engineering blueprints" containing all design steps and editable elements, whereas AI is currently better at drawing "final effect images" (like a single picture) and still struggles to directly create and package such intricately structured "blueprint files." Manus provided a detailed explanation of how this blueprint could be designed, which can be considered a completion of the task. However, the final presentation effect was still somewhat semi-finished and might require multiple rounds of revisions.
Figure: Manus's output webpage effect
Lovart's delivery result is shown in the figure below. The TikTok marketing plan was presented as a visual graphic, and a dynamically displayable webpage was generated. Similarly, no downloadable Figma or Photoshop source files were generated, but multiple images were generated, covering from background to subject, which can be re-edited using the top toolbar, or downloaded for secondary editing with software like PS. The video also showed the dynamic effect of putting stars into bubbles. The overall plan has a relatively high degree of completion.
Promotion plan webpage:
Flowith's final deliverable also lacked Figma/PSD source file download links. In terms of final visual presentation, it had a strong GPT-4o feel. The video generation had some deviations, failing to show a connection with the beverage, and not quite matching the main visual tone and style from before.
Text-based color scheme:
Primary color: Cosmic Deep Blue #0A1F3C Symbolizes the deep and vast starry sky, establishing the brand's technological and dreamy tone. Suitable for large backgrounds and main brand text, creating a calm and mysterious visual effect.
Accent color: Bubble Silver #C0C0C0 Represents the vibrancy and sheen of bubbles, emphasizing the drink's vitality and fresh taste. Suitable for bubbles in the logo, auxiliary graphic elements, and for use with Cosmic Deep Blue to add layers and a fashionable feel.
Highlight color: Grapefruit Yellow #FFD700 Extracts the color of grapefruit, symbolizing fresh and natural fruit flavor, injecting vitality into the overall color scheme. Suitable for grapefruit elements in the logo, highlighting important information, and pairing with deep blue and silver to form a visual focal point.
Figure: Flowith generated promotional video
Scenario 5: In-depth Research Scenarios: Lovart Absent, Flowith and Manus Each Have Strengths
In in-depth research scenarios, an Agent's core competitiveness lies in a set of interconnected "long-chain" capabilities: breaking down vague topics into executable steps, precisely retrieving and integrating information from vast data, providing verifiable sources for conclusions, and self-correcting during execution.
Only when these capabilities form a closed loop can an Agent handle complex research tasks lasting tens of minutes and involving hundreds of operations. Therefore, powerful Agent products must integrate six core capabilities—long-range planning, ultra-long context understanding, enhanced retrieval, tool invocation, self-reflection, and credible traceability—to build a reliable and scalable system, ensuring outputs in complex research scenarios are accurate, efficient, and verifiable.
Since Lovart is an Agent focused on design capabilities, it is not suitable for this scenario. We focused on evaluating Manus and Flowith.
Below is a test case:
Prompt:
Research why Claude 4 can continuously code for 7 hours, and why Flowith can achieve ultra-long context + continuous self-generation. What are the main technical principles behind this? How is this achieved when base models still have context limitations? What is the significance of long-range autonomy of large models for humans? The long-duration working capability of intelligent agents has never before been used as an evaluation metric for models or AI products, why is it being continuously mentioned now? Research relevant papers on websites like arXiv, are there any related discussions? From social media and institutional media, observe how industry KOLs express themselves. Output a complete report: including interpretation of basic principles, practical significance for the industry, current bottlenecks, and future development route judgment.
Flowith and Manus took approximately 5 minutes each. Flowith generated a final report of 12,375 words, and 32 node documents.
Manus generated a final report of 12,813 words, and 9 documents covering segmented topics, such as technical background research.
From the quality of the final in-depth reports generated, both Agents are hard to distinguish. However, their different working processes and principles allow us to use them in different scenarios.
Both Flowith and Manus can integrate "retrieve first, then reason, then execute" into a single pipeline, but their implementation focuses differ: Flowith relies on "Oracle Mode + Knowledge Garden" to consolidate large amounts of raw data into visual nodes, then completes deep analysis through thousands of steps in an infinite chain;
Figure: Flowith's work interface
Manus emphasizes a "browser/terminal/code editor" three-in-one tool orchestration, aiming to directly transform research results into deliverable products.
Figure: Manus's work interface
When research materials are vast and require multi-person collaboration and frequent iteration, Flowith's "Knowledge Garden + Visual Chain" is more suitable.
If we prioritize quick output, with results needing to be finalized as websites, scripts, or finished documents, and data compliance risks are manageable, Manus's tool orchestration would be more convenient.
If the scenario requires both ultra-long context and local or private cloud deployment, neither current solution is complete enough, and both require a hybrid autonomous retrieval framework or a specialized research agent as a complement.
Concluding Remarks
Are Users Willing to Pay for Agents?
Lovart is still in its invite-only free trial phase; Flowith sets its entry-level subscription at $19.9/month; Manus, after opening registration on May 13, priced its basic version at $19/month, matching Flowith. Converted at current exchange rates, this price point is approximately 1700 RMB per year.
For different user groups, the same price means fundamentally different decision-making logics:
General users are mainly driven by light experience and interest, and will only consider moving from free to paid when a certain function significantly improves personal efficiency.
Professional users (content creators, freelancers) regard high-quality output and stability as essential. If a tool can save production time and improve delivery standards, a monthly fee of around $20 is relatively acceptable.
B2B users (teams, enterprises) prioritize security and compliance, permission management, and API/workflow integration; as long as the product is reliable in these aspects, the monthly fee is not an obstacle to payment.
Therefore, the decisive turning point may not be a few percentage points improvement in model performance, but whether the product can convert "curiosity traffic" into "monthly repurchase GMV" through clear efficiency dividends. Only when individuals or small teams are willing to turn their monthly "coffee money" into "Agent efficiency fees" can these Agent products truly cross the commercialization threshold.
Figure: Manus Pricing
Figure: Flowith Pricing
ima Knowledge Base
The AI Energy Station compiles basic popular science and tutorials on AI application practices, covering fundamental theories, technical research, value alignment theories, and industry development reports from leading global companies, top scientists, researchers, and market institutions in artificial intelligence, as well as global AI regulatory policies. It helps AI novices get started and advanced players track the latest AI knowledge.
Recommended Reading
Google 2025 I/O Conference: Returning to the Smart Glasses Battlefield, Revolutionizing Itself with AI
Microsoft Released 50 New Things Overnight, Building an "Eden" of Interconnected Agents
Text-to-Image! Hunyuan Image 2.0 Brings AI Image Generation into the "Zero-Latency" Era