My thoughts on the products of Button Space and Manus

background

These are some temporary notes for the newly released Button Space. Similarly, these are only for the upgrade description of the AIP multi-agent platform. They are temporary notes with a slightly colloquial tone.

A previous manuscript titled " My Further Upgrade and Implementation Paradigm for Multi-Agent Platforms " has just published a version of the paradigm theory. If Manus is an introduction (not yet mature), the emergence of Button Space has set the tone for the idea of ​​multi-agent collaboration, clarified the future direction and determination of multi-agent collaboration, and also reflected its form. At present, it is very likely (just possible) that it will be the form of Agent in the next 2-4 years.

Overview

The acceleration is to avoid excessive exploratory design pre-research, product form thinking, and technical pre-research. For AIP, the design part is the most time-consuming part. The following is the overall architecture design 2 years ago:

Obviously, 25 years later, this architecture is somewhat outdated. It used to be a single-agent architecture, but now it needs to be adjusted to a multi-agent collaborative architecture.

A relatively clear industry development route is needed as a guide to build and improve the following contents:

• Product form: clarify the design of the interactive form from product input to output results

• Toolkit: Extraction or specification of toolkit and integration with large model (MCP)

• Data assets: Industry Agent or general Agent and data assets

  
  

The technical solution and technical route are also relatively clear, which is also the original positioning. It is simple and easy to integrate with its own business and enterprise-level scenarios. It is small and can be private.

Every designer has his own opinions and ideas, and I have my own thoughts.

Clarify the content

Once the ideas and directions are determined, we will soon be able to improve the form of usable products and optimize the usage scenarios.

Product form

The scenario needs to be closed. The closed loop form of the task scenario is the goals and results of the input and output. For example, in the document writing scenario, writing a romance novel is the goal, and outputting the novel is the result. This forms a closed loop. Publishing the novel is another scenario, and each scenario is implemented one by one.

The combination of multiple Agents and Agents from different industries, when entering a scene, selects different Agents for different scenes to resolve the difference between general Agents and business Agents.

The scene interaction form is based on planning, execution, real-time output, and result integration. The final form is a route to form a scene processing idea. This is indeed a reference to Manus's interaction design. The following are two examples. The real-time Agent execution output is on the right.

Example of multi-agent data analysis:

Example of multi-agent document review:

The above is the embodiment of the multi-agent collaborative product form. There is still a lot of room for optimization in this form in the future, and the path is relatively clear. At present, we will consider landing it first.

Toolkit Integration

The toolkit is a pain point in the early stages of agent development. Some of the problems have been solved through API calls, and the original interface architecture has also been processed. The previous positioning was to provide interface capabilities in the data asset module, but it was always felt that it was not the optimal solution.

The emergence of MCP seems to be able to solve this problem, but there are uncertainties and the maturity of the technical framework, so we still need to wait and see whether it can become an industry standard. At present, under the promotion and use of many large manufacturers, the combination idea of ​​MCP has been clearly determined, but services are needed, and the client framework needs to customize the combination method to solve the problem of separation between the intelligent platform and the tool platform. There are also standardization issues. The specific implementation of MCP is just another way.

The combination of tools will soon be able to unify the early-stage Agent and the toolkits of all parties into a standard (this standard specification is what is needed), which can make the engineering relationship and the overall design clearer, and is also conducive to market promotion. It is expected that the construction and adaptation of the MCP service will be completed in about 2 months.

Data asset integration

By distinguishing between intelligent agents and industry expert agents, we can see whether they are valuable. When the model capabilities are similar, the general results will appear to be readable but not usable.

In other words, it has little to do with industry business or the business itself. The main focus here is on direct result delivery.

The combination of data assets and Agent landing is as follows:

• General Agent: These are visible or commonly used logical thinking, such as the current large models.

• Business Agent: This is business-oriented and can be as large as years of company data or as small as a table or an Excel data.

  
  

This will include processes such as data labeling, scoring, collection, cleaning, output, training, etc. The current goal is to provide Agent scenarios. For example, reports will be output by data analysis agents, and work and Sunday reports will be output from work agents. These will form new forms of interaction and output, and form clear specifications (which are also needed), thereby forming engineering and systematization.

Summarize

The above are some references for learning and studying the application scenarios and designs of large models. My energy and ideas are limited, and I hope that students who are interested can communicate with each other.