Function Compute:Create an agent with Quick Creation (no-code)

Step 1: Go to the creation page

1. Go to the Agent Runtime Console. If you are using it for the first time, you must complete the Service-Linked Role (SLR) authorization. In the dialog box that appears, click Confirm.

2. On the Agent Runtime page, click Create Agent.

3. Select Quick Start to open the visual configuration interface.

Step 2: Select a model

Assign a name to your agent and select an underlying language model.

1. Enter a custom Agent name.

2. Select a model: In the Select model section, choose a model from the Large language model dropdown list.

   • Quick start: If this is your first time, click Configure Qwen model with one click. The system guides you to use an API Key from Alibaba Cloud Model Studio to quickly deploy a default model (qwen-max). You can go to the Model Studio Console to activate the service for free and obtain an API Key.

   • Add other models: Click Add model and refer to Model management for instructions. AgentRun supports multiple model integration methods:

     • API model: Directly call external model services, such as Alibaba Cloud Model Studio, Qwen, or third-party commercial models, with an API Key. This method is flexible and ideal for integrating existing external model services.

     • One-click deployment: Deploy and host your own open-source or self-trained models. This method enables data privacy and model control, and is suitable for scenarios that require model customization.

• Model parameter configuration: Adjust model inference behavior by configuring parameters such as inference mode, temperature, and token limits.

  Note

  Recommendation: A lower temperature value results in more stable and deterministic model responses. A higher value leads to more creative and diverse responses. For scenarios requiring precise answers, set a lower temperature (for example, 0.1 to 0.3). For creative tasks, set a higher temperature (for example, 0.7 to 0.9).

  To simplify the process in Quick Creation mode, some advanced parameters such as top_k and seed are unavailable. For fine-grained parameter tuning, convert the agent to the pro-code mode.

Step 3: Write and optimize prompts

A prompt is the instruction you give to an agent, guiding its actions, role, and the rules it must follow.

1. Write a system prompt: Define the agent's core responsibilities and workflow. You can use the preset Prompt or write a custom one in the interface.

2. Optimize the prompt (optional): After writing a prompt, use the AI Prompt Assistant to refine its phrasing or structure for improved clarity and effectiveness.

3. Configure prompt variables: Use variables to enable the agent to handle dynamic information. Define variables in the prompt by using the {variable_name} format, such as {user_name}.

   Example prompt:

   ## Workflow
   1. Understand the customer's question.
   2. Query order information for the customer `{user_name}`.
   3. Provide an accurate and helpful reply.

Step 4: Configure tools or sandboxes

Tools and sandboxes extend an agent's capabilities, allowing it to interact with the external world or execute code. Examples:

• Configure the web scraping tool to let the agent search for and process content from web pages, and convert HTML to Markdown for easier use.

• Configure the BrowserTool sandbox to provide the agent with a browser instance for automated testing and debugging.

• Configure the Code Interpreter sandbox to enable the agent to execute Python code, manage a file system, and process data files in a secure environment.

To configure tools or sandboxes:

1. Create tools and sandboxes: If no tools are available, go to Resource configuration > Tool market, search for the tool you need, click Install, and follow the instructions in the console. You can also refer to Tools and skills and Sandbox services for more information.

2. Select a tool or sandbox.

Step 5: Configure access credentials

Configure credentials, such as an API Key, for the agent's endpoint to protect it from unauthorized calls. Credentials are automatically managed and injected by AgentRun's credential management service.

1. In the Access credential section, click Inbound: Access credential.

2. Select a credential mode:

   • No credential (not recommended): The agent's endpoint can be accessed anonymously from the public internet, which poses a security risk. This mode is for functional testing only and must not be used in production environments.

   • Use existing credential (recommended): To ensure the security of your agent, select this option. If you do not have a credential, click and refer to Credential management to create one.

Step 6: Finalize and test

1. After you complete the preceding steps, click Create Agent in the upper-right corner to save the configuration and create the agent.

2. After the agent is created, you can enter test messages in the test area to interact with the agent and verify its functionality.

Step 7: Publish a version and perform canary release

AgentRun supports version management and canary releases. After you modify the agent's prompt, tools, or model, first publish a version, then create an endpoint and enable a secondary version (canary release) to split a small portion of traffic to the new version. After confirming that the new version is stable and reliable, gradually increase the traffic percentage until the release is complete.

1. In the left-side navigation pane, choose Versioning and canary release.

2. Publish the current version: Click Publish Version, enter a Version Description that describes the main changes and features, and then click Publish Version.

3. Create an endpoint:

   1. Enter an Endpoint name. From the Primary version dropdown list, select the version number you just published.

   2. Enable the secondary version (canary release): Select the Enable secondary version (canary release) checkbox, select a secondary version, and configure the traffic splitting percentage for the primary and secondary versions.

Run a comparison test

You can use the Comparison Test feature to compare the performance of different models or versions with the same input. The results are displayed side-by-side for easy comparison, which helps you adjust your model, prompt, or tool configurations.

1. In the left-side navigation pane, choose Configuration and debugging.

2. In the Model comparison test section, click Comparison test.

3. Select a Test type and the objects to compare:

   • Model test: Select different models to compare their response quality and performance.

   • Endpoint test: Select different endpoints to compare the responses that result from different prompts, configurations, and tools.

4. After you make your selections, enter the same question to see the test results displayed side-by-side, as shown in the following figure:

Convert to pro-code mode

When your business logic becomes more complex and requires finer-grained control, you can convert your agent from Quick Creation to pro-code mode with a single click. The converted code has a clear structure, which allows you to perform further development without starting from scratch.

• Code editing: Directly edit the agent's code implementation for greater flexibility and control.

• Version control: Use version control to track the history of every code change.

To convert the agent:

1. On the Configuration and debugging page of the agent, click Convert to code mode in the upper-right corner.

2. Carefully read the information in the dialog box, and then click Convert to code creation. If you only want to view the agent's code, click Code preview to download and view the code locally.

   Important

   This conversion is a one-way operation and cannot be undone. You will need to configure the agent through code. For more information, see Create an agent by using code (pro-code).

3. After the conversion is complete, the agent is switched from the no-code visual interface to the pro-code mode.

Integrate into an application

In the left-side Integration and publishing section, refer to Agent integration and publishing to quickly integrate your agent into your front-end web pages, back-end applications, and more. It supports UI integration, code integration, and ecosystem integration.

Runtime configuration

Resource configuration

Network configuration

Network mode: PUBLIC, which allows the default network card to access the public internet.

Q1: Using Quick Creation agents in production

A:

Yes, but it is recommended only for scenarios with simple logic and controllable traffic. For production environments handling complex business processes or requiring high availability, we strongly recommend converting to pro-code mode after you complete prototype validation. This provides greater control, stability, and maintainability.

Q2: Runtime, scaling, and performance

A:

The agent runs in an instance environment based on Function Compute (FC). This environment automatically scales based on the number of sessions and requests, so you do not need to manually configure the number of instances. The cold start latency typically ranges from milliseconds to seconds, and performance depends on multiple factors, such as model response time and tool invocation latency. You can view detailed performance metrics on the agent's Observability page.