What's Behind the Manus Hype: How Can Agentic AI Products Build Lasting Competitive Advantages?

🚥 Koji's Note: Next week, Crossing will release the third episode of its "20 Questions" series. Three weeks after Manus went viral, we plan to use 20 questions to unpack AI Agents — one of the hottest topics of 2025.

One of this episode's guests is Yan Wang, the author of this article.

Subscribe to the Crossing podcast on Xiaoyuzhou and Apple Podcast to catch next week's episode.

🚥 About the Author: Yan Wang

Applied Scientist at Samsara, PhD from Columbia University, formerly at Pinterest and Microsoft, with nearly 40 papers published in top international conferences and journals. Personal website: Computing Life[1]

Manus launched recently and quickly took over the Chinese internet. After using it extensively, I find the product genuinely inspiring. It grasps something crucial in the Agentic AI race: compound effects. This article explores what really matters for products like Deep Research, Cursor, or Manus over the long run — where the competitive battlegrounds lie, what builds real moats, and what doesn't.

Before diving into the three key dimensions of Agentic AI competition, let me explain why I find Manus so impressive. Contrary to what many influencers suggest, Manus didn't emerge from nowhere. It wasn't unprecedented. Rather, it follows a clear lineage.

Two product forms are most directly relevant:

1. Agentic research products like Gemini, Perplexity, and OpenAI's Deep Research. You input a topic or request, and they research across the web to produce detailed, substantive reports.
2. Agentic generation products like Cursor, Devin, or Gamma. You give them a prompt, and they write code, draft documents, or create slides — delivering finished output directly.

In 2024, both categories matured significantly, crossed the usability threshold, and saw viral growth.

But one major pain point remained: you could either do research, or you could write code. The two weren't effectively connected.

This is actually a subtle but critical problem. The whole point of Agentic AI is that through self-iteration and autonomous decision-making, it helps us complete complex tasks end-to-end. But if in practice I have to constantly think, "I'll use OpenAI's Deep Research for this part, then copy-paste the results into Cursor to generate some visualizations, then merge everything and throw it into Gamma for slides" — this completely defeats the purpose of Agentic AI and undermines its very reason for being.

What makes Manus stunning is that it connects the entire pipeline.

On one hand, it can conduct research in agentic fashion, browsing the internet to gather comprehensive, detailed material.

On the other hand, it can analyze and visualize that material to generate final deliverables — websites, illustrated reports, or presentations. This end-to-end application was nearly impossible to achieve with previous products.

Add to that Manus's polished execution and high degree of completeness. A product with precise vision and solid usability naturally becomes a hit.

The Compound Effect of Tools

But these observations don't really get at the more essential characteristics, or advantages, of Agentic AI.

In my view, one defining feature of Agentic AI compared to traditional AI is its compound effects across multiple dimensions.

In Manus's case, a key reason for its success is this: it can use more tools than previous products. This is actually a non-trivial distinction.

In Agentic AI products, the jump from 6 to 8 tools improves the user experience far more than the jump from 2 to 4. That's because AI tools can combine and reinforce each other. An AI that only writes code and searches text won't gain much from adding image search.

But if it can also write reports and make presentations, then adding image search suddenly makes its output far more vivid and compelling — meaningfully advancing the product experience.

This is precisely what Manus did. Setting aside its other innovations, even just combining Deep Research and Cursor into one product — this simple increase in tool count — immediately enabled scenarios that previous products couldn't achieve.

This is the first compound effect of Agentic AI.

As we increase the number of callable tools, the benefits are combinatorial and explosive. This is a very practical way to enhance user experience, but it doesn't build an effective moat. The reason: with tools like Cursor readily available, simply wiring up a tool for AI to autonomously call is quite straightforward.

Setting aside product quality, replicating a Manus isn't inherently difficult. And trying to build a moat purely by racing to add more tools isn't a sustainable strategy.

The Compound Effect of Data

Agentic AI has similar compound effects in other dimensions too.

One often overlooked aspect is data.

This doesn't mean LLM pretraining data — the "I used 2T tokens, you used 3T, so I'm better" kind. In the Agentic AI era, it carries deeper meaning. Specifically, it's not about data volume per se, but about data acquisition, organization, and externalization across the entire lifecycle. When working with humans, we often appreciate that "with an elder at home, you have a treasure." When a factory machine breaks, the veteran worker knows exactly where to tap it back to life. An experienced doctor can diagnose with a few touches, while fresh graduates need batteries of tests to reach similar conclusions.

This is a classic example of data-driven advantage. For humans, two main things happen in this process: first, experience accumulation — whether the mechanic or the doctor, they've encountered similar failures or cases over decades. On this foundation comes the second step: synthesis and organization, i.e., knowledge structuring. At this point, the knowledge is internalized in their memory, which suffices for humans. But since AI-human communication currently relies entirely on written exchange, an additional step of knowledge externalization is usually needed — distilling it into explicit documentation that AI can use.

Therefore, for Agentic AI, maintaining a cycle of knowledge accumulation, organization, and externalization is crucial.

Consider a software engineering example we mentioned earlier. If you give a coding AI a 100,000-line codebase with some tasks, its probability of nailing everything in one shot isn't high.

But if you give it time to gradually read, comprehend, digest, and categorize the code — then summarize and distill what it learns into documents — its coding work becomes much easier.

Here[2] is an example of such a document. It covers the codebase's basic architecture, design philosophy, and which functions live in which files. For older projects, we can add historical context. With this documentation support, spatially, the AI knows how to precisely locate code to modify instead of blindly creating new files and rewriting everything from scratch. Temporally, the AI also knows what attempts were made before and what the current design approach is, avoiding the "ghost wall" phenomenon of circling back to previous designs.

Thus, the data we're discussing isn't mere accumulation or token stacking, but a long-term, automatic or semi-automatic process of accumulation, comprehension, and distillation. For a given client, the longer the AI works with them, the more such knowledge it accumulates. Compared to a newcomer with no background — even if the latter is more intelligent — users will find the former more comfortable, more attuned to them. This secondary processing of knowledge systems is an effective moat.

Similarly, this data accumulation has its own combinatorial compound effects. With more historical data and synthesized documents, the AI can form deeper insights through comparison and reflection.

To some extent, this is a process of transforming traditional knowledge systems into AI-friendly ones.

AI-friendliness isn't a binary state, but something requiring time to precipitate and develop. I'd even compare it to co-evolution between humans and nature.

On one hand, AI mines, refines, and accumulates from raw knowledge bases.

On the other hand, as users work with AI, they increasingly realize that making data easily accessible to AI greatly benefits their own work. So they become more willing to adapt their workflows to AI's data management processes. This brings additional benefits — for instance, tribal knowledge previously lost in Zoom meetings can now be captured and deposited into document libraries via Zoom AI Companion, available for AI to leverage on the user's behalf.

This creates a mutually reinforcing positive loop. And this mutual adaptation and tacit understanding is a very strong moat.

The Compound Effect of Intelligence

Agentic AI has another fascinating characteristic: intelligence itself compounds. This may be less intuitive than tools or data. But a tool's intelligence level affects the Agentic user experience from multiple angles.

At the most basic level, smarter tools better understand how to efficiently grasp user needs and how to combine a few tools for maximum gain.

A less intelligent LLM might flail about, calling many tools without gathering sufficient information. A more thoughtful LLM appears more methodical, solving problems quickly through clever combination of just a few tools.

Another relevant factor: if you've compared Gemini and OpenAI's Deep Research, you'll find they're completely different leagues.

Gemini feels more like mechanically following predetermined instructions: brainstorm keywords, search, autonomously decide which pages to scrape, then summarize based on those pages.

OpenAI's Deep Research feels more proactive, with stronger self-iteration capabilities. It first formulates a plan, then searches with different keywords according to that plan. During searching, it may dynamically adjust strategy based on results. The final output is often genuinely insightful, not just answering the question but autonomously extending to valuable next research directions. The gains from this autonomy are non-linear.

Given that few companies currently have the capability for autonomous LLM R&D, and that LLM training is intensely resource- and capital-intensive, this to some degree constitutes a meaningful moat.

The Battlegrounds

But I want to emphasize: these three compound effects aren't parallel additive relationships — they have multiplicative effects that spark and amplify each other.

More tools create more outlets for information processing and accumulation, leaving data interfaces across project management, search, document generation, and other dimensions for AI to learn from. As AI analyzes this information, it continuously improves its reasoning and comprehension. This co-evolutionary effect is quite evident in Manus.

Initially, Deep Research could conduct deep research, and Cursor could write code and documents — but once fused within a single Agentic platform, the AI could perform further logical processing, orchestration, and publishing on the research findings.

With information, tools, and intelligence stimulating each other in the same closed loop, users get a smoother, smarter end-to-end experience.

Therefore, the key to Agentic AI competition may lie in how quickly you can expand to the right side of these exponential growth curves for tools, data, or intelligence. Because spending equivalent effort to grow tool count or data volume yields relatively limited returns early on. But once it crosses a tipping point, the power of exponential growth becomes apparent.

On the right side of the curve, each additional tool or bit of data makes the user experience noticeably smoother. This is likely where Agentic AI products should compete and build moats. Of course, this exponential curve won't extend infinitely — it may follow an S-curve. Beyond some point, adding new tools or information becomes constrained by system complexity and resource investment, causing growth to slow and potentially hit bottlenecks. At that stage, deeper architectural and collaboration mechanism innovations may be needed to sustain co-evolution.

From the above discussion, we can also see that building moats around tools isn't particularly reliable. Building moats around LLM intelligence requires massive resources, while building moats around data may be the most accessible and feasible approach.

Beyond data accumulation itself, what's perhaps more important is the process and methodology of how to accumulate. Because data can be copied, but systematically externalizing tacit knowledge, structuring沉淀, and managing data efficiently — this is extremely difficult to replicate.

This resembles corporate culture. Once a strong methodology and process system for data management and knowledge externalization is established, even if competitors copy the data and tools, they can hardly replicate this implicit organizational capability in the short term.

Thus, in the long-term competition of Agentic AI products, the hardest thing to crack isn't the scale of data or intelligence, but this systematic organizational capability around data and tool usage.

In Summary

Starting from Manus, we've seen where the competitive keys and moats lie in Agentic AI.

But more importantly, this competition isn't merely about比拼 tool counts or data scale — it's about how organizations adapt to the profound transformation of the AI era. The future winners may not be those with the strongest technology, but those who truly understand how AI and humans co-evolve and can establish sustained, stable collaboration mechanisms.

This may be the real revelation that Agentic AI brings us.

---

P.S. At the end of this article, a recommendation for Yan Wang's AI course FROM USERS TO BUILDERS: Evolve Yourself to Embrace the AI Era [3] — interested readers are welcome to check it out.

References [1] Computing Life: https://yage.ai/

[2] Here: https://github.com/grapeot/web_agentic_ai/blob/master/claude-tooling/app/README.md

[3] FROM USERS TO BUILDERS: Evolve Yourself to Embrace the AI Era: https://www.superlinear.academy/c/ai/