Yesterday, at midnight, Monica.im released Manus, the world's first general-purpose AI Agent product. If DeepSeek marked a stunning comeback for Chinese companies in the large-scale model arena, then Manus represents a leap forward in the Agent domain.
Manus has surpassed OpenAI's DeepResearch in the GAIA benchmark (a standard evaluation system for general AI assistant capabilities), claiming the top spot. Unlike traditional AI assistants, Manus stands out for its generalization across diverse tasks and its ability to autonomously execute and deliver final results.
In simple terms, Manus can tackle a wide range of complex and variable tasks—writing research reports, planning trips, analyzing financial statements, and more—while directly delivering complete outcomes.
There's plenty of buzz around Manus right now. Whether it truly achieves the status of a "general-purpose Agent" remains up for debate, but the unique product philosophy behind it offers valuable lessons.
For instance, Manus champions "Less structure, more intelligence", advocating for fewer structural constraints on AI and relying instead on the model's autonomous evolution rather than pre-designed workflows.
Previously, I translated an article titled "Painful Lessons for AI Entrepreneurs: Betting on Model Accuracy is a Product Trap; Leveraging Model Flexibility is the Answer". In it, the author, after studying over 100 AI startup projects, argued:
AI products shouldn't expend too much energy on a model's limitations. Instead, they should capitalize on the autonomy and flexibility of large-scale models. As models continue to evolve, the added value of software will diminish, but the boundaries of application will expand significantly.
This perspective finds validation in Manus: through robust engineering, they've integrated existing functionalities into a seamless experience, delivering smoother usability and superior results.
Beyond "Less structure, more intelligence," according to the WeChat public account Automate.AI, the Manus team shared four additional insights at a closed-door meeting today:
First, while Manus is positioned as a "general-purpose Agent," its primary focus remains on information gathering and research.
From the official use cases, it covers obvious workplace scenarios as well as lifestyle applications like trip planning—spanning research report writing, data analysis, lead generation, lifestyle tasks (e.g., travel planning), and education.
This approach starkly contrasts with Zhipu's AutoGLM. AutoGLM leans toward being a personal life assistant, helping users send red packets, order takeout, book rides, or check routes.
This highlights a divergence in philosophy: Manus squeezes the model's potential to tackle complex tasks, while AutoGLM starts with simple daily tasks to create a product accessible to the masses.
Compared to AI sending red packets, workplace scenarios like copywriting are far more suited to current AI capabilities. These tasks are time-intensive, have rigid demand, and align with large models' strengths—given sufficient context and environment, they outperform humans by a wide margin.
Second, Manus fully displays all steps of its execution process and mimics human usage habits.
In demo cases, Manus breaks down user tasks into smaller, actionable steps, fetches the necessary information, allows users to pause workflows and provide feedback, and then resumes. It also saves all gathered data, synthesizing it into the user-requested report at the end. This mirrors DeepSeek's chain-of-thought display, showing users how results are derived, fostering greater trust.
What's more, it was noticed that Manus simulates human behavior in demos—flipping through PDFs page by page or browsing websites one by one. Theoretically, large models can instantly process vast amounts of data for greater efficiency.
The reason for this design? The current internet is built around human habits. For compatibility and generality, Manus mimics these patterns. As AI capabilities advance, Agents may evolve toward more efficient collaboration models.
Manus' true strength lies in two areas: generalization across tasks and autonomous execution with final delivery.
Functionally, each of Manus' features has precedents—like Deep Research, Artifact, or Web Search. What sets it apart is its use of large models' reasoning and generalization to integrate these functions, eliminating the need for users to juggle multiple tools while delivering a seamless, effective experience. That's where it shines.
This universality stems not just from the team's deep engineering expertise but also from their unique understanding of AI products.
Central to this is their mantra: "Less structure, more intelligence"—reducing structural limits on AI and leveraging the model's evolving capabilities over manually preset processes.
In short, it's about minimizing restrictions and harnessing the model's growth to tackle more tasks efficiently.
This reminds me of a recent insight from Lukas Petersson, co-founder of Andon Labs (YC W24 batch). After studying over 100 YC alumni projects, he concluded:
Many AI products over-invest in current model limitations, but in the long run, startups should bet on opportunities that maximize model autonomy and flexibility.
Prompt engineering can improve AI performance, but its ceiling is low. A better strategy? Wait for stronger models. In the interim, products with greater autonomy will outperform.
This is playing out now. At the meeting, the Manus team revealed that its performance already surpasses roughly three-quarters of Agent startups from YC W25.
Manus embodies this philosophy—they believe AI can outsmart humans and give models ample room for autonomous planning and execution.
Granted, its use cases don't yet break new ground, but as model capabilities grow, application boundaries may expand over time.
To this end, Manus introduced the Agentic Hours per User (AHPU) metric to measure time efficiency in delegated tasks, aiming to boost productivity via parallelism.
The team also shared that Manus' single-task cost is $2—well below industry averages—with room for further optimization.
Beyond these points, per Automate.AI, the Manus team shared additional insights at the closed-door session: