Skip to main content
Understanding agent-based LLM applications is hard. Agentic systems are often designed to make autonomous decisions and chain together dozens—sometimes hundreds or even thousands—of operations. Each operation results in one or more spans. This level of complexity makes it nearly impossible to trace an agent’s behavior by manually inspecting individual spans or traces.
Image

​
Why Agent Visualization Matters

Instead of wading through hundreds of spans, stakeholders like developers and AI PMs want to quickly answer questions like:
  • What are the most common execution paths?
  • Where are things failing?
  • Which agents are calling which?
  • Is there any self-looping behavior?
These kinds of questions are not easily answered by trace lists alone. That’s why we built Agent Graph and Path Visualization—a way to abstract individual spans into node-based graphs. These visualizations help map your application flow in a human-understandable way, bridging the gap between granular telemetry and system-level insight.

​
Implementing Agent Visualization for Arize

Agent and node visualization is designed to track high-level workflow components and their relationships, not every individual operation. Think of it as a “logical flow map” rather than a detailed trace view. This is powered by span metadata that identifies agents and defines the transitions between them. We automatically track these attributes for popular frameworks, so no additional implementation is needed. For other frameworks, custom agents, or hybrid systems, you can use custom implementation.

​
Frameworks with Built-In Support

The following frameworks have built-in support for agent metadata through their auto-instrumentors:
  • Automatically tracks agent nodes and graph transitions
  • Uses native metadata.langgraph_node and metadata.langgraph_step
  • Handles agent metadata and state transitions
  • No additional implementation needed
  • Automatically tracks graph.node.id and graph.node.parent_id
  • Handles agent handoffs through _handoffs tracking
  • No additional implementation needed
  • Automatically tracks agent roles and task relationships
  • Includes agent metadata in spans
  • No additional implementation needed
  • Handles agent metadata via OpenInferenceTracingProcessor
  • Tracks handoffs between agents
  • No additional implementation needed
  • Tracks agent names and team relationships
  • Includes agent metadata in spans
  • No additional implementation needed

​
Custom Implementation

​
When Custom Implementation Is Needed

Custom agent metadata tracking is required when:
  1. Using frameworks without built-in support:
    • Vanilla OpenAI / Anthropic calls
    • Custom agent implementations
    • LangChain without agent components
    • Other unsupported frameworks
  2. Using hybrid instrumentation:
    • Mixing auto-instrumented frameworks with custom code
    • Building custom agents that interact with instrumented frameworks

​
Required Metadata Attributes

To enable agent and node visualization, include the following metadata:
AttributeDescriptionRequired
graph.node.idUnique name for the agent/node✅
graph.node.parent_idID of the parent node (if applicable)Optional, but recommended. We will infer the graph using the relationship of the spans within the trace if this is not included

​
Example Hierarchy

AttributeExample
graph.node.id"input_parser", "research_agent"
graph.node.parent_id"main_orchestrator"
The example creates this structure: 
main_orchestrator
├── input_parser
├── content_generator
│   ├── research_agent
│   └── writer_agent
└── quality_checker
You do not need to annotate every span. Only annotate those components you want represented in the graph.

​
Example Implementation

​
Basic Pattern

# 1. Add graph attributes to your spans
def add_graph_attributes(self, span, node_id: str, parent_id: str = None):
    span.set_attribute("graph.node.id", node_id)
    if parent_id:
        span.set_attribute("graph.node.parent_id", parent_id)
    span.set_attribute("graph.node.display_name", node_id.replace("_", " ").title())

# 2. Use in your tracing code
with self.tracer.start_as_current_span("my_operation") as span:
    self.add_graph_attributes(span, "my_component", "parent_component")
    # Your logic here...

​
Multi-Level Hierarchy

# Root level (no parent)
with tracer.start_as_current_span("main_workflow") as main_span:
    add_graph_attributes(main_span, "orchestrator")
    
    # Child level
    with tracer.start_as_current_span("parse_input") as parse_span:
        add_graph_attributes(parse_span, "parser", "orchestrator")
        
        # Grandchild level
        with tracer.start_as_current_span("validate_input") as validate_span:
            add_graph_attributes(validate_span, "validator", "parser")