Agent Development

Function Overview

The agent development platform provides a one-stop solution for building agents. It offers a zero-code orchestration tool to help developers quickly create AI applications and uses canvas-based node design to cope with complex service scenarios. The platform provides functions such as role setting, plug-in extension, and workflow orchestration. It supports knowledge base management, RAG retrieval, and intelligent prompt optimization to ensure accurate and reliable interaction. It also provides short-term dialog memory and long-term personalized storage, and can be published through various channels such as APIs and web pages, helping developers efficiently build professional agent applications.

• Orchestration capability

  Table 1 Agent platform orchestration capability

  Function	Description
  Orchestration Mode	Allows users to call and create agents through conversational interaction, enabling service personnel to build native AI applications within just five minutes without writing any code. Provides flexible canvas-based node orchestration, enabling agents to accurately solve complex service scenario problems. Provides a multi-agent orchestration canvas to combine multiple expert agents. The system automatically schedules and allocates agents to enable the most appropriate agent to solve problems.
  Model selection	The platform provides multiple third-party models. Third-party deep thinking models, such as DeepSeek, have been adapted to the agent development platform.

• Role instructions

  The agent development platform can set human-like features through role instructions to improve the interaction realism. The platform presets role instruction templates. You can also enable intelligent addition to allow the large model to output a better role prompt. Role setting helps developers create highly human-like voice interaction scenarios.

• Skills

  The core capabilities of agents are derived from its skill system. Developers can continuously expand the function scope of models by integrating plug-ins and designing workflows.

  Table 2 Agent platform skills

  Function	Description
  Plugin	You can seamlessly connect to various platforms and services through APIs to quickly expand the functions of an agent. The platform provides a variety of built-in plug-ins that are ready to use. You can also customize plug-ins and encapsulate any API into a tool for flexible calling.
  Workflow	Workflows are visual tools for building complex function logic. By flexibly combining multiple task nodes, you can design multi-step automated processes, significantly enhancing the ability of agents to handle complex tasks.

• Knowledge base

  It provides out-of-the-box enterprise RAG (management, testing, and retrieval policy configuration).

  Table 3 Knowledge center on the agent platform

  Function	Description
  Managing knowledge bases	The list displays knowledge base names, IDs, descriptions, and creation timestamps. You can view, delete, edit, and publish knowledge bases.
  Creating a knowledge base	You can enter the knowledge base name, description, and knowledge base upload type, and upload a document. You can select built-in knowledge (in .txt, .docx, .pdf, .doc, or .xlsx format) and reference knowledge (linking to an existing knowledge base).
  Uploading knowledge base documents	You can view the document list, import new documents, and view and delete knowledge bases.
  Slice policy	It supports custom policy configuration for text slicing, including sentence slicing character identifiers, maximum slicing length, and maximum word overlap ratio.
  Slice details	You can view basic document information, slices, and slice knowledge details.

• Prompt engineering

  The platform provides prompt templates and development tools so that anyone can develop high-quality prompts with ease.

  Table 4 Prompt functions

  Function	Description
  Prompt writing	Brainstorm large batches of prompts and select the best ones as candidates. Example prompts can be imported. Model configuration is supported. Variables in prompts can be defined. Outcome preview is supported. Historical records can be viewed.
  Prompt comparison	Candidate prompts can be compared (prompt difference comparison and outcome comparison).
  Prompt evaluation	You can create an evaluation case set, evaluate the prompt quality using multiple evaluation methods, and select the optimal prompt.
  Prompt optimization	The prompt engineering platform provides an automatic prompt optimization function. Leveraging heuristic algorithms for self-optimization and gradient optimization for selecting the best prompts, it automatically refines existing prompts based on specific evaluation cases.
  Automatic prompt generation	With the intelligent template matching and layout optimization technologies, the system automatically generates a high-quality prompt template based on the content entered by users and the original prompt template and model capabilities.
  Prompt application	Instruction configuration and the LLM node in the workflow allow you to save and reference prompt templates.

• Memory capability
  The dialog-based workflow memory capability of the agent development platform supports short-term variables to maintain dialog continuity, supports long-term memory for personalized, in-depth interactions, and allows users to optimize parameter configurations to strike a balance between performance and effectiveness.

  Table 5 Memory capability of the agent development platform

  Function	Description
  Variable	Stores temporary session data (such as the current user input and intermediate calculations), and is valid only for a single conversation.
  Memory parameter	Determines the number of previous dialog turns to retain as context.
  Long-term memory	Persistently stores users' personalized data (such as preferences and historical behavior), encodes content into vectors, and supports semantic search.

• Dialog experience

  The dialog experience of agent development supports full-process visualization and quick fault locating and configuration optimization through debugging.

  Table 6 Dialog experience of the agent development platform

  Function	Description
  Agent Logo	Brand logo of the agent, which is used for visual identification and generally reflects the function or personality of the agent.
  Prologue	Initial greeting for the interaction between the agent and user, which is used to set the tone and guide the user. The greeting can be customized by the user.
  Recommended questions	Typical questions preset at the beginning of each dialog, helping users quickly understand the capability scope of the agent. You can customize the questions.
  Follow-up questions	Follow-up questions are proactively raised during the conversation with the agent to clarify requirements or deepen interaction.
  Preview Debugging	Real-time test and optimization tool for agent functions. You can view the execution result and call details.

• Global configuration

  Dialog-based workflows now support a global configuration entry, which enables configuration of the dialog experience, default model, global feature switches, and workflow definitions.

  1. You can configure prologues or recommendation questions in dialog-based workflows, and these can be automatically generated.
  2. The default model can be configured to serve as the default option for newly added nodes. A one-click setting allows quick modification of the global model, enhancing configuration efficiency.
  3. The long-term memory switch and memory variable definitions can be customized.
• MCP service

  The platform tool calling supports the MCP protocol. Developers can quickly enhance agent functionality by integrating the MCP service.

  Table 7 MCP service description

  Function	Description
  MCP Market	Multiple preset MCP services, such as AutoNavi Map, Ticket Query Tool, and Bing Search, are available on the platform. You can integrate and call them with one click.
  Custom MCP services	The platform allows developers to customize MCP services and quickly create them based on the MCP service address.

• Release channel

  Agents can be released in different modes and channels, and can be released as APIs or to web pages. In this way, agents can be quickly implemented in actual services, balancing technical integration and user experience.

  Table 8 Agent platform release channel

  Function	Description
  API	Encapsulates the agent into standard APIs for developers to directly invoke. This mode applies to scenarios where in-depth integration with other systems (such as enterprise internal applications and third-party services) is required.
  Application library	The management account releases the agent to the built-in application library of the platform so that platform users can quickly call the application.

Why Agent Development Platform?

• Comprehensive plug-in integration capabilities

  The agent development platform continuously expands the capability boundaries of agents through a rich plug-in ecosystem. The platform presets dozens of practical plug-ins, such as file processing and code interpreter, for out-of-the-box use. It also provides open capabilities for creating custom plug-ins. Developers can encapsulate internal APIs, third-party tools, or self-developed functions into plug-ins and enable agents to invoke the plug-ins with simple parameter configuration.

• Flexible knowledge base management

  The platform provides an intelligent knowledge base system, which allows developers to build dedicated data asset libraries. Local documents (PDF, Doc, and PPT formats), web page data, and structured databases can be quickly imported to the knowledge base. The agent automatically learns the data and creates indexes to implement accurate knowledge retrieval and content generation. For example, after an industry report is uploaded, the agent can automatically generate a competitive product analysis report.

• Modular workflow design

  The platform uses the zero-code and low-code visualized development mode, allowing developers to quickly build complex task flows by dragging and dropping nodes. It provides nearly 100 standard components, such as natural language processing, logic judgment, and code execution, to meet the development requirements from simple Q&A to multi-step tasks (for example, building an intelligent customer service workflow that includes requirement analysis, solution design, and cost accounting).

• Full-link agent development support

  The platform provides a complete development and debugging system, covering code compilation, plug-in testing, and workflow verification. Developers can observe each step of the agent execution in real time, view the context memory, plug-in calling result, and inference process, and quickly locate problems through hierarchical log analysis (for example, tracing the dialog logic and data calling path when the intelligent customer service handles user complaints).

• Abundant MCP expansion capabilities

  Multiple preset MCP services, such as AutoNavi Map, Ticket Query Tool, and Bing Search, are available on the platform. You can integrate and call them with one click. The platform allows developers to customize MCP services and quickly create them based on the MCP service address.

Platform Architecture

The agent development platform provides one-stop AI application building capabilities. The platform provides multiple capability combinations to support Agent development scenarios with different complexities.

Basic general capabilities

The agent development platform can be used by multiple tenants and managed in a unified manner. Session management and historical dialog records can be viewed. Agents can be published as APIs.

Core functional modules

• Workstation: The Workstation module allows you to create zero-code applications, orchestrate applications through workflows, create custom plug-ins through plug-in management, and access applications.
• App treasure box: The app treasure box contains the AppGallery, workflow market, and external apps, and supports various out-of-the-box apps.
• Plug-in market: More than 10 types of plug-ins are preset in the plug-in market, such as the file processing plug-in and code interpreter, which are out-of-the-box.
• MCP market: Multiple external MCP APIs are preset for users to easily invoke.
• Knowledge base: The knowledge base can be uploaded and managed, and the slicing policy of the knowledge base can be set.
• Prompt engineering: Prompt templates can be created, compared, evaluated, and optimized.

Application scenarios

Agents can be used in various scenarios:

• Knowledge engineering scenarios: Knowledge bases are mounted to apply an agent to professional fields such as finance and healthcare.
• Dynamic planning scenarios: The agent development platform is used in scenarios such as chatbot and tool calling.
• Process orchestration scenarios: The agent development platform is used in complex service scenarios through the component library and workflow orchestration.
• Multi-agent orchestration scenarios: Agents can be used for precise intent distribution.
• Innovation exploration scenarios: Agents can be used in scenarios such as embodied AI.

Highlight 1: Zero-Code Creation of Agents

• Intelligent prompt generation: The platform provides an intelligent prompt generation function. Even if you do not have a foundation in large models, you can quickly generate suitable prompts to easily guide agents to complete tasks.
• Rich skill configuration: The platform provides various skill configuration options, covering knowledge bases, plug-ins, and workflows. Adding skills is as simple as building blocks. You only need to click and select skills to enable the agent to master new skills, such as accessing knowledge bases in specific domains to become experts in these domains.
• Quick creation and running: The entire creation process is very simple. You can bring an agent online in a few steps. From conception to actual application, the time is greatly shortened, allowing zero-code users to quickly have their own intelligent assistants.

Highlight 2: Low-Code, Flexible Creation of Complex Workflows

• Visualized workflow design: The GUI encapsulates various functions into nodes, such as the large model node and intent recognition node. Developers can drag and connect nodes like a puzzle to build complex workflows. The logic is clear and visible.
• Code enhancement flexibility: Code can be written between nodes, providing more customization space for developers with programming experience. For example, you can add several lines of data processing code to the code node to implement special data conversion and meet personalized requirements.