What Will Manus Bring to the Web3 DeFAI Scene?

By Haotian

I woke up to many friends urging me to check out #manus, touted as a truly universal AI Agent with the ability to think independently, plan and execute complex tasks, and deliver complete results. Sounds cool—but beyond the waves of anxiety in social media about job displacement, what does it mean for the next big breakout in web3 DeFi? Here are my thoughts:

1) Roughly a month ago, OpenAI launched Operator, a similar product enabling AI to autonomously perform tasks in a browser—like booking restaurants, shopping, ticket reservations, or food delivery. Users can visually monitor progress and take over control at any time.

Yet this Agent received little discussion, mainly because it's driven by a single model operating within the traditional tool-calling framework. As soon as users realize they still need to intervene for critical decisions, trust in autonomous execution evaporates.

2) On the surface, manus appears similar but expands into more application scenarios such as resume screening, stock research, and property purchasing. The real difference lies beneath—the underlying architecture and execution system. Manus is powered by multimodal large models and innovatively employs a multi-signature system.

In short, to mimic human execution of the PDCA (Plan-Do-Check-Act) cycle, multiple large models collaborate, each specializing in a specific phase. This reduces decision-making risks from any single model while improving efficiency. The so-called "multi-signature system" is essentially a decision-validation mechanism for multi-model collaboration, ensuring reliability by requiring consensus across several specialized models before actions are executed.

3) In comparison, manus clearly stands out. Combined with its compelling demo videos showcasing seamless operations, it delivers an impressive user experience. Objectively speaking, however, manus’ innovation over Operator is just the beginning—not yet a revolutionary leap.

The key lies in the complexity of tasks it can execute reliably, and how well the system handles unstructured, non-standard user prompts—specifically, its fault tolerance and success rate in delivering correct outputs. Otherwise, could we immediately apply this innovation to mature DeFi use cases in web3? Clearly not.

For example: In a DeFi scenario, if an Agent must make trading decisions, it requires an Oracle-layer Agent responsible for collecting, validating, and analyzing on-chain data, while continuously monitoring prices to identify trading opportunities. This poses significant real-time analytical challenges—an arbitrage opportunity valid one second ago may vanish by the time the Oracle’s large model transmits data to the trading execution Agent (arbitrage window closure).

This reveals the biggest weakness of such multimodal large models when making execution decisions: How to connect to live networks, access real-time on-chain data, analyze trading opportunities instantly, and capture trades accordingly. In web2 environments like e-commerce, where order prices don’t fluctuate rapidly, maintaining dynamic balance across multimodal systems is manageable. But in blockchain environments, this challenge is constant and relentless.

4) Overall, manus will undoubtedly trigger anxiety in web2 circles—many repetitive clerical and information-processing roles may indeed be replaced by AI. Let them worry.

But when assessing its actual impact on advancing DeFi applications in web3, we must remain objective:

We must acknowledge: Its significance is substantial. The concept of an LLM OS, the philosophy of "Less Structure, More Intelligence," and especially the multi-signature system offer valuable inspiration for integrating DeFi and AI in web3.

It corrects a major misconception among most DeFi projects—don’t assume that a single large model can immediately achieve full autonomy in AI Agents, including independent thinking and decision-making. That’s simply unrealistic in financial contexts.

Realizing true DeFi requires solving far more complex issues: overcoming individual AI model capability limits, ensuring atomicity in multimodal collaborative interactions, unified resource scheduling and allocation across multimodal systems, and robust fault-tolerance and error-handling mechanisms.

For instance:

• An Oracle-layer Agent collects and analyzes on-chain data, monitors price movements, and provides reliable data feeds;
• A Decision-layer Agent evaluates the Oracle-provided data, performs risk assessment, and formulates actionable strategies;
• An Execution-layer Agent carries out selected plans based on real-world conditions, optimizing gas fees, managing cross-chain states, resolving transaction ordering conflicts, etc.

Only when all these Agents are equally powerful and integrated within a comprehensive system framework can a genuine DeFi revolution truly begin.