Multi-task Parallelism

Core Capabilities

• Multi-role collaboration: Multiple specialized agents perform their respective duties, simulating the division of labor and cooperation of human developers.
• Full-process coverage: Requirement analysis and code delivery are automatically connected to streamline an end-to-end workflow.
• Intelligent collaboration: Team cooperation breaks the limitations of single-point intelligence, boosting development efficiency and code quality.

Agent Team Architecture

An Agent Team typically consists of the following components.

Generating Code Using Agent Team

1. Switch to the agent team mode.
   1. Access CodeArts Agent Space by referring to Enabling CodeArts Agent Space.
   2. Click Coding.
   3. At the bottom of the input box, select AgentTeam to switch to the agent team mode. The selected model is displayed on the right. You can select a model as required from the drop-down list.

2. Explain what you need. Type your task into the box, and click or press Enter to submit your request.

   Create a mini-program for tracking sleep quality. First analyze existing products in the domestic market and summarize their pros and cons, and then develop a mini-program with distinctive features.

3. The team leader creates tasks and assigns them to teammates. Once the team leader receives the task instructions, they will deconstruct the main objective into smaller, more specific sub-tasks. Based on the nature of these subtasks, the team leader dynamically creates and assigns teammates, while simultaneously generating a task list. The task list contains parallel or serial sub-tasks (such as Task 1, Task 2, Task 3... Task N).
   Figure 1 Breaking down and assigning tasks
   
   • Switching to the team visualization mode: Click on the top menu bar of the chat window. You can switch from the chain of thought (CoT) mode to the team visualization mode. This mode intuitively displays the upstream and downstream relationships of task assignment, allowing you to clearly identify the dependency structure of tasks and understand the data flow and execution sequence between each node. When you want to focus on the work progress of a specific sub-agent, simply hover your mouse over the node to quickly view its real-time working status via a floating pop-up window. Click in the pop-up window to go to the CoT view of the subtask.
     Figure 2 Team visualization mode
     
   • Opening the task overview: Click on the top menu bar of the chat window. Under Task Overview, you can view the task duration, number of completed tasks, and number of sub-agents.
     Figure 3 Task overview
     
   • Viewing the task list: Click on the top menu bar of the chat window. Under Task List, you can view the current tasks.
     Figure 4 Task list
     
     

     : That the task is complete.

     : The task is in progress.

     : The task is not started.

   • Viewing task deliverables: Click on the top menu bar of the chat window. Under Task Product, you can view the final task deliverables. Click the generated file to view its content in Documents or Editor.
     

     Once the task execution is completed, you need to first accept and confirm the output files. Then the files are displayed here.

   • Viewing context information: Click on the top menu bar of the chat window. You can view details about the skills and MCP tools invoked by the current chat under Context Information.
   • Viewing context usage: Click on the top menu bar of the chat window. You can view the context usage (percentage of used tokens) of the current chat under Context usage in real time. This helps you intuitively manage your current context usage and prevent disruptions to the session continuity due to exceeding limits.

4. Teammates collaborate to execute the tasks. When the team leader assigns sub-tasks (e.g., Task 1) in the task list to corresponding teammates, it dynamically configures an identity/role for each teammate and generates a specialized prompt to ensure the teammate accurately understands and executes that specific task. Upon receiving their task, teammates will further break it down into a list of specific to-dos. Based on the requirements of these to-dos, each teammate invokes corresponding tools to enhance their capabilities and complete complex operations.

   If you need to view the execution details of a subtask, click the target task item to automatically jump to the subtask page.

   Figure 5 Viewing the subtask execution progress
   

5. View and confirm the results. After each subtask is completed, the execution results are reported to the team leader. The team leader then centrally collects the outputs, verifies them, and consolidates the pieces to generate a final summary report.
   Figure 6 Subtask execution results reported to the team leader
   
   Figure 7 Final result confirmation
   
   • Viewing and processing generated files in the editor: View the content of generated files under Editor. Click Keep (Ctrl+Y) on a segment to accept the generated content, or click Undo (Ctrl+N) to reject the generated content. Alternatively, at the bottom of the file, click Keep All (Ctrl+Enter) or Undo All (Ctrl+Delete) to accept or reject the generated file, respectively. You can also directly modify the content.
     

     If the editor is not displayed properly, click (Toggle Tool Panel) in the upper right corner of the top menu to open the tool panel. On the displayed panel, click Editor.

   • Viewing and processing generated files in the temporary storage area: Move the cursor to the row where the target file is located, click to accept the file, or click to reject the file. If there are multiple files, you can click Accept All to accept all files or Reject All to reject all files.

Creating a Team Template

In CodeArts Agent Space, you can run the /save-team command to encapsulate roles and scenes from completed tasks into a reusable template. When creating new tasks later, the system can directly invoke this template without having to replan the architecture from scratch.

1. Refer to Generating Code Using Agent Team to complete task execution.
2. Click in the input box, select Shortcuts, and choose /save-team from the displayed menu.
3. Enter the name of the team to be created (for example, new-dev-team) and click .

   The syntax for the command is /save-team Team name, where you can customize Team name. If you need to update an existing team template, simply run /save-team Name of the team to be updated.

4. View the generated role and scene documents.

   After the files are generated, you can view them under Editor > EXPLORER.

   • Role documents are stored in the .codeartsdoer/agent-team/roles directory under the project folder.
   • Scene documents are stored in the .codeartsdoer/agent-team/scene directory under the project folder.
   Figure 8 Viewing generated role and scene documents
   

5. Create a new task based on the generated role and scene documents.

   Enter the following prompt in the input box (with the team name specified in 3) and click .

   Use new-dev-team to create a children's sleep-aid mini-program.

   After receiving the prompt, the AI will begin constructing the task based on the roles and scenes defined in the specified template.

Obtaining Agent Team Logs

When encountering exceptions or errors during code development in the agent team mode, refer to the following to retrieve logs for troubleshooting. The following operations apply only to Windows and macOS.

Navigate to the local paths %USERPROFILE%/AppData/Local/Temp/agent-team-log and view relevant logs.

Figure 9 Viewing the agent team logs