1. Planning Phase (Planning Agent)

• Task Decomposition:
  • Parse task objectives based on user input (such as natural language requirements, data characteristics, goals, etc.).
  • Decompose complex tasks into fine-grained subtasks to form a clear subtask list.
• Define Task Chain:
  • Organize subtasks into a task chain (pipeline) according to their dependencies.
  • If user requirements are clear, directly generate a standard task chain that meets the needs; otherwise, generate one through task reasoning.

2. Tool Registration Phase (Tool Register)

• Register Predefined Tools:
  • Predefined tool library: includes operators for data sampling, cleaning, evaluation, generation, etc.
  • Dynamic tool generation: automatically register new tools by package scanning and operator description extraction (e.g., via get_desc).
  • Tool description vectorization: generate semantic representations of tools by combining functional descriptions, input/output types, sample data, etc., for subsequent retrieval and recommendation.

3. Task Dispatch Phase (Task Dispatcher)

• Assign Tasks to Agent:
  • Distribute tasks to the corresponding agent sequentially or in parallel according to the task chain.
  • Provide context tracking for the task chain (e.g., using sessionKEY for multi-turn dialogue tracking).

4. Task Execution Phase (Execution Agent)

• Generate Tools:
  • Dynamically write the required code for tasks (such as Python scripts) and debug the code.
  • During execution, automatically obtain task parameters (such as user input or the output of the previous task) and fill them into the code.
  • When parameters or context are missing, the Execution Agent will prompt the user for completion.
• Process Tasks:
  • Execute node operations within the task chain, including invoking predefined tools or dynamically generated tools.
  • Pass execution results in real-time to subsequent task nodes or return them to the user for review.

5. Result Evaluation Phase (Evaluation Agent)

• Debug and Validate:
  • Perform syntax, logic, and data integrity checks on task execution results.
  • For code-based tasks, check whether the code execution results meet expectations (e.g., returning results in JSON format).

6. Analysis and Summary Phase (Analysis Agent)

• Summarize and Generate Report:
  • Analyze the task execution process and results, generating highly readable reports.
  • Generate contextual summaries for multi-turn dialogues, helping users quickly review the overall execution of the task chain.