Summary
As agentic workflows scale, inspecting and debugging multi-agent delegation topology (SequentialAgent, ParallelAgent, LoopAgent, tools, memory) via terminal logs or raw Python code becomes challenging.
We propose adding a built-in CLI tool adk graph (and Web UI interface) that provides:
- Reverse Inspection / AST Parser (Python → Graph): Reads existing ADK Python agent files and renders their multi-agent topology visually.
- Visual Builder & Code Generator (Graph → Python): Drag-and-drop agent visual editor (integrating concepts from ADKGraph).
- Live Execution & Debugging Visualizer: Highlights active agents and tools on the visual graph in real time as events stream over SSE/WebSockets during runner execution.
User Experience / Proposed CLI
# 1. Inspect & debug an existing agent script
adk graph path/to/agent.py
# 2. Launch the visual agent editor to construct a new agent system
adk graph --new
When run, adk graph will:
- Spin up a lightweight local inspection server.
- Open an interactive web UI in the browser showing the agent topology.
- Allow real-time step execution with visual active-node highlighting.
Key Features
- Bi-directional Editing: Changes made in code update the graph visual; canvas edits compile directly back to production-ready
adk-python code.
- Hierarchical Orchestrator Containers: Support nested visual group containers for
SequentialAgent, ParallelAgent, and LoopAgent.
- Live Event Debugger: Visual node glow/state animation during active agent execution, tool calls, and LLM reasoning steps.
Proposed Architectural Changes
- CLI extension: Add
graph subcommand in google.adk.cli.
- Inspector engine (
google.adk.cli.graph): Dynamic agent object visitor & Python AST static analysis module.
- Web UI bundle: Embedded lightweight React visual canvas server.
Additional Context
This proposal is based on the open-source ADKGraph visual IDE built specifically for Google ADK. We are willing to contribute and maintain this feature via a Pull Request.
Is this feature aligned with the ADK roadmap? We welcome maintainer feedback before submitting our PR implementation!
Summary
As agentic workflows scale, inspecting and debugging multi-agent delegation topology (
SequentialAgent,ParallelAgent,LoopAgent, tools, memory) via terminal logs or raw Python code becomes challenging.We propose adding a built-in CLI tool
adk graph(and Web UI interface) that provides:User Experience / Proposed CLI
When run,
adk graphwill:Key Features
adk-pythoncode.SequentialAgent,ParallelAgent, andLoopAgent.Proposed Architectural Changes
graphsubcommand ingoogle.adk.cli.google.adk.cli.graph): Dynamic agent object visitor & Python AST static analysis module.Additional Context
This proposal is based on the open-source ADKGraph visual IDE built specifically for Google ADK. We are willing to contribute and maintain this feature via a Pull Request.
Is this feature aligned with the ADK roadmap? We welcome maintainer feedback before submitting our PR implementation!