Overview

MultiAgent is a powerful multi-agent system designed to coordinate multiple agents working together to accomplish complex tasks. The creation and configuration process of MultiAgent is streamlined and efficient, utilizing a Planner-Executor-Reviewer approach for task execution. Organizations can customize their AI agent teams based on different business scenarios, significantly enhancing the efficiency and flexibility of AI applications within the enterprise. The platform comes with various pre-configured AI agents, including AI Coder IDE, Browser Use, and Computer Use. These agents are ready to use out of the box without complex configuration.

The operational mechanism of MultiAgent is as follows:

flowchart LR

 subgraph P["Planner"]
    direction TB
        P1["Task Clarification
        Identify and clarify user instruction intent"]
        P2["Planner
        Generate high-quality Task Todo List"]
        P3["Action
        Dynamically generate Action List for Tasks"]
  end
 subgraph R["Runner"]
    direction TB
        T(["Task-1"])
        T2(["Task-2"])
        T3(["Task..."])
    direction LR
        A3["Action1"]
        A4["Action2"]
        A5["Action3"]
        A6["Action4"]
  end
 subgraph Rev["Reviewer"]
    direction TB
        RV["Enterprise Custom Review Standards"]
        D1{"Human
        Takeover Needed?"}
        D2{"Full Task
        Update Required?"}
        D3{"Pending Tasks
        Need Update?"}
        D4{"Current Task
        Need Re-execution?"}
        End(["Continue Task Execution"])
  end
    H2["Human User Takeover"]
    H1["User Issues Task"]
    P1 --> P2
    P2 --> P3
    P -- Serial --> T & T2 & T3
    T -- Parallel --> A3 & A4
    A6 -- Task Execution Complete --> RV
    RV -- Not Pass --> D4
    D4 -- N --> D3
    D4 -. Y .-> T
    D3 -. Y .-> P
    D3 -- N --> D2
    D2 -- N --> D1
    D2 -. Y .-> P
    D1 -- Y --> H2
    D1 -- N --> End
    A4 -- Serial --> A5
    A5 -- Serial --> A6
    RV -- Pass --> NT["Start Next Task"]
    H1 -- Submit Task Instruction --> P

     H1:::human
     H2:::human
     P1:::planner
     P2:::planner
     P3:::planner
     T:::runner
     T2:::runner
     T3:::runner
     A3:::runner
     A3:::Peach
     A4:::runner
     A4:::Peach
     A5:::runner
     A5:::Peach
     A6:::runner
     A6:::Peach
     RV:::reviewer
     D1:::decision
     D2:::decision
     D3:::decision
     D4:::decision
     End:::reviewer
     NT:::reviewer
    classDef human fill:#e3f2fd,stroke:#90caf9,stroke-width:2px
    classDef planner fill:#e8f5e9,stroke:#81c784,stroke-width:2px
    classDef runner fill:#fffde7,stroke:#ffe082,stroke-width:2px
    classDef reviewer fill:#ede7f6,stroke:#b39ddb,stroke-width:2px
    classDef decision fill:#fce4ec,stroke:#ec407a,stroke-width:2px,stroke-dasharray:4 2
    classDef Peach stroke-width:1px, stroke-dasharray:none, stroke:#FBB35A, fill:#FFEFDB, color:#8F632D

Key Features

• Custom AI Agent Team Building: The Multi-Agent platform enables enterprises to flexibly assemble AI Agent teams based on specific business needs, greatly enhancing the flexibility of AI applications within organizations.
• Rich Pre-configured AI Agents: The platform comes with various built-in, ready-to-use AI Agents, such as AI Coder IDE, Browser Use, and Computer Use.
• Multi-scenario Role Support: Provides AI Agent roles covering multiple domains including development, product, testing, algorithms, data, and marketing, meeting the needs of different enterprise departments.
• Wide Application Scope: Multi-Agent can be applied to various business scenarios such as in-depth research, data insight analysis, bid document generation, and financial account detail generation.
• Efficient and User-friendly: Simple operation that greatly improves work efficiency. For example, users can assemble an AI data analysis Agents team within minutes.

Creation and Configuration

Creating a MultiAgent system is straightforward, similar to setting up a single Agent. Key aspects include:

• Global Settings: By default, global settings are collapsed. Here you can configure memory, input/output capabilities, and other modules such as knowledge base, database, logs, and insights, consistent with single Agent functionality.
• Node Addition: Users can add various Agents to the MultiAgent canvas by clicking "Add Node". Available Agent types include:
• Pre-configured Agents: Ready-to-use Agents built by GPTBots with predefined capabilities, such as Online Search and Browser Use.
• Template Agents: Agents requiring user configuration to define their roles and capabilities.
• Agent Integration: To build a MultiAgent team, simply select the desired Agents, add them to the canvas, and connect them with lines. This links their capabilities for collaborative task execution.

Agent Introduction

Planner Agent

The "Planner" is the core Agent of MultiAgent, responsible for requirement clarification, task planning, execution task generation, and task completion quality review. It's recommended to choose the most capable model version, though selecting an inference model may significantly increase runtime.
The Planner supports custom task control mechanisms, such as dynamic task adjustment, user takeover options, and setting maximum retry limits for individual tasks to avoid excessive token consumption.
The Planner can also be configured with memory, tools, knowledge base, and database capabilities for enhanced functionality.

Computer Use Agent

Computer Use Agent supports custom identity prompts and configurable maximum task iteration limits to prevent excessive token consumption. It also supports user takeover functionality, allowing human users to take control and remotely operate the Computer when the Agent cannot properly advance tasks.
Computer Use runs in a sandboxed Linux environment and can utilize computer apps, file systems, and system commands to complete tasks.

Browser Use Agent

Runs in a sandboxed browser environment for browser-based task execution. Has low LLM capability requirements and can be used regardless of whether the LLM supports image recognition.

Coder IDE Agent

Completes tasks through code generation, running in a sandboxed CLI environment. Excels at searching, executing APIs, generating web pages, and creating documentation.

Online Search Agent

Online Search possesses autonomous capabilities for retrieving and obtaining URL page content, making it particularly useful in information network detection scenarios.

Execution Process

The MultiAgent task completion process follows these steps:

1. Task Input: Users input task descriptions and submit. The Planner determines whether to execute directly or request clarification.
2. Task List Generation: Once instructions are clear, the Planner generates a task list for user confirmation before automatic execution.
3. Dynamic Execution: During execution, the Planner dynamically generates Action Lists for each task. After each task completion, the Planner evaluates quality and decides whether to adjust or update the task list.
4. Task Completion: Upon completion of all tasks, MultiAgent summarizes and provides final results. Users can ask follow-up questions to update task outcomes if needed.

Best Practices

• Capability Description: Ensure accurate and detailed capability descriptions for each Agent to help the Planner achieve efficient task allocation.
• Model Optimization: Choose models for the Planner that balance performance and speed, avoiding delays from excessive reasoning.
• Control Mechanisms: Set appropriate iteration limits and user takeover options to optimize resource consumption and task reliability.