MCP or Not, Manus Made a Choice – A Deep Dive into AI Agent Architectures
By Kelvin Lu
First Contact with Manus
Manus has made waves as China’s second major AI innovation following DeepSeek in 2025. At first glance, it appears to be a typical multi-agent system, but beneath the surface, it introduces a paradigm shift in AI development.
To evaluate its capabilities, I tested Manus with various real-world scenarios. When asked to compare salaries and job opportunities for developers and engineers, it autonomously planned a workflow, browsed the web for information, extracted relevant data, and compiled a structured report. The results were impressively comprehensive, though not flawless.
Next, I fed it a deliberately flawed query—asking it to compare housing prices across different suburbs, including a typo and a fictitious location. Manus not only identified the typo but also flagged the fake suburb. The end result? A detailed report with well-organized charts. Although not perfectly refined, its performance was remarkable.
One key difference was Manus’ transparency. Before executing any task, it set up a sandbox environment, offering an explicit breakdown of its process—unlike other AI vendors who keep such operations obscured.
Previously, someone claimed to have jailbroken Manus. The developers, however, were unfazed. Their response? A nonchalant shrug and: “That’s how it’s designed.”
Furthermore, Manus openly detailed its internal structure—a departure from the norm where most AI systems disclose only minimal functionality. Its distinct architecture challenges the dominance of the MCP protocol in the AI ecosystem.
What Is MCP?
MCP has become a trending topic in AI circles. It allows AI applications to integrate with a standardized network of agents, enabling seamless interoperability. Several MCP servers and service catalogs have surfaced despite the concept being relatively new.
“MCP is an open protocol that standardizes how applications provide context to LLMs. Think of MCP like a USB-C port for AI applications—offering a universal way to connect models to diverse data sources and tools.”
— Anthropic
Despite MCP’s surging popularity, many misunderstand its mechanics. Let’s break it down:
Core Components of MCP
- Clients: Interfaces such as Claude or code editors like Cursor that interact with MCP servers, converting user queries into function calls.
- Servers: Back-end services that expose tools, resources, and prompts that AI models can access.
- MCP Protocol: The framework governing communication between clients and servers.
Why Is it Called MCP?
The name “MCP” has sparked confusion, with some mistakenly assuming it stands for “Memory, Context, and Personalization.”
According to official documentation, MCP is not about conversation history but rather focuses on structuring communication between AI clients and servers. Anthropic has even stated that future iterations of MCP aim for stateless operation, meaning MCP servers will not retain conversation logs.
Introducing CodeAct – A Rival Approach
Most AI agent frameworks either rely on template prompting or instruct LLMs to “think step by step.” However, XingYao Wang’s research paper introduced a novel concept: reasoning through code instead of plain text or JSON.
This discovery inspired Manus to explore a new methodology, diverging from MCP. While MCP structures service interoperability, Manus optimizes AI reasoning through its own variant of CodeAct.
MCP vs. CodeAct: A Comparative Study
The two primary challenges of AI agents include:
- Reliability: Ensuring consistent, accurate performance.
- Extensibility: Allowing for seamless updates and adaptability.
Manus and Anthropic took different paths to tackle these issues. While MCP improves feature discoverability for AI applications, Manus prioritizes reliability via centralized planning. By integrating CodeAct, Manus potentially achieves superior stability and efficiency.
There are inherent limitations to MCP:
- MCP applications may integrate smart features more easily, but they risk broken workflows due to decentralized planning.
- MCP does not define how individual servers process requests—it only standardizes interaction.
- Despite its vision for universal AI connectivity, MCP currently supports only local connections, as authentication mechanisms remain unresolved.
- Many MCP servers serve as mere wrappers around existing functions, adding minimal innovation.
Meanwhile, Manus sidesteps these pitfalls with a structured, centralized planning approach.
Final Thoughts
AI agent development is evolving rapidly, with multiple frameworks competing for dominance. While MCP addresses cross-service interoperability, its role in reliability and error mitigation remains questionable.
The technology landscape changes at breakneck speed. Once-revolutionary concepts—like early RAG models—quickly become outdated as new ideas emerge. MCP is still in its infancy and has yet to prove its long-term viability.
Before we champion MCP as the definitive standard, let’s remember past framework overhauls that failed despite initial hype:
- CORBA
- Enterprise Java Beans
- Jini
- DCOM
- Web Service Discovery Protocol
Only time will tell if MCP will avoid joining that list.
