Step 1: Activate the product and create a workspace
Complete identity authentication by using your Tencent Cloud account.
When first entering the Agent Development Platform, the system will automatically create one "Enterprise" and one default workspace for the primary account.
Step 2: Create an application and select a mode
Within the workspace, create an "application", which typically corresponds to a business scenario (such as after-sales intelligent customer service).
Select the appropriate application mode based on business complexity.
|
Standard Mode | Suitable for lightweight tasks that "start with a simple agent and expand capabilities on demand," such as beginning with basic Q&A and gradually adding capabilities like knowledge base queries, tool invocation, or plugin actions. You can first launch a prototype with a minimal feature set, then iteratively supplement more knowledge and capabilities based on business needs. These enhancements are seamlessly integrated by the system for continuous performance improvement. For details, see Creating a "Marketing Copy Generation Assistant" in Standard Mode. |
Multi-Agent Mode | Suitable for complex tasks where "a single agent cannot complete the work and requires team collaboration," such as conducting comprehensive analysis of a batch of data to form decision recommendations. You can configure dedicated agents for different stages like "problem analysis, data research, computation, risk control verification, and conclusion drafting." The system automatically coordinates division of labor in the background to deliver a result after multiple rounds of verification. For details, see Creating a "Stand-up Comedy Content Creation Assistant" in Multi-Agent Mode. |
single-workflow mode | Suitable for standardized business processes that have been solidified, such as ticket routing, after-sales processing, and approval workflows. You can drag and drop steps like "determine whether conditions are met → call APIs to query data → provide handling suggestions based on results" to orchestrate a reusable workflow. This allows agents to repeatedly execute the same "assembly line," stably and efficiently handling repetitive tasks. For details, see Creating a "Web Content Summarization Assistant" in Single-Workflow Mode. |
For detailed description, see the following documents:
Step 3: Configure the knowledge base and plugins
Bind one or more knowledge bases to the application, import business documents or Q&A content as the source for what the agent "knows." Enable plug-ins based on business needs, for example:
Internet search
Internal system API
Code execution, and so on
For detailed description, see the following documents:
Step 4: Build the agent logic
For example, configure prompts, model parameters, and response style within the application, optionally chaining simple workflows. Complex scenario examples:
Use "workflow orchestration" to build multi-step processes and integrate external systems, databases, and so on.
Combine nodes such as large models, knowledge search, conditional judgments, API calls, and code execution in workflows.
Examples of multi-domain scenarios:
Using a Multi-Agent architecture, the master agent orchestrates multiple expert sub-agents, each responsible for different subtasks.
For detailed description, see the following documents:
Step 5: Dialogue debugging and effect evaluation
Debug conversations within the console to observe whether the responses are accurate and stable. Write test cases for key scenarios and conduct batch evaluations through evaluation capabilities:
Knowledge search effectiveness (recall rate, accuracy, and so on)
Conversation quality (relevance, completeness, fluency, and so on)
Step 6: Publish the Application and Configure Channel Access
Publish the application from the test environment to the production environment, and select the access method based on business requirements. For example:
Embed components in websites or backend systems to connect to ADP.
Access WeCom, WeChat Official Accounts, Mini Programs, LINE, Telegram, and so on
Self-developed frontend via API/SDK calls.
For detailed description, see the following documents:
Step 7: Monitor Operations and Continuously Optimize
View metrics such as request volume, success rate, response latency, and Token consumption for the application on the platform, analyze common issues and satisfaction levels, and continuously optimize:
Supplement or cleanse the knowledge base
Adjust workflows and prompts.
Optimize model selection and routing policy.