Agent Graphs
Agent graphs in Langfuse provide a visual representation of complex AI agent workflows, helping you understand and debug multi-step reasoning processes and agent interactions.
Example trace with agent graph view (public link)
Get Started
The graph view is currently in beta, please feel free to share feedback.
There are two ways a graph appears for a trace:
- Inferred from observations. Have an observation with any observation type except
span,event, orgenerationin your trace. Langfuse then interprets the trace as agentic and shows a graph, inferred automatically from the observations' timings and nesting. - From the LangGraph integration. When you use the LangGraph integration, the graph shows automatically.
Observation Types: See all available Observation Types and how to set them. LangGraph: See the LangGraph integration guide for an end-to-end example on how to natively integrate LangGraph with Langfuse for LLM Agent tracing.
View modes: Aggregated and Expanded
The same trace can be drawn two ways. Switch between them with the Aggregated / Expanded toggle in the top-left corner of the graph; your choice is remembered across traces.
The distinction is how repeated steps are treated — one node per step name vs. one node per call:
| Aggregated (default) | Expanded ("as it ran") | |
|---|---|---|
| A node is… | one unique step name | one individual call |
| Repeated calls | collapse into a single node with a counter | appear as separate nodes |
| Loops / cycles | drawn as cycles (edges loop back) | unrolled into an acyclic graph (DAG) |
| Reads like | the agent's overall shape | the exact run, step by step |
| Best for | grasping structure & complexity at a glance | following or debugging a specific execution |
Aggregated
The Aggregated view is a compact overview. Steps that share a name are merged into a single node with a counter — retrieve_docs (3/3) means that step ran three times — and a loop that calls the same tool repeatedly is drawn as a cycle rather than a long chain. This keeps even a busy agent readable: you see how many distinct steps there are and how they connect, without every individual call on screen. It's the fastest way to answer "what does this agent do, and how complex is it?"
Expanded
The Expanded view shows the full process as it ran. Every call is its own node, so three litellm_request calls are three distinct nodes, and loops unroll into a directed acyclic graph in execution order. It's more literal and closer to the trace tree — the right view when you want to follow a specific run end to end or pin down exactly where something happened. Larger traces naturally produce more nodes here than in the aggregated view.
Neither is "correct" — they answer different questions. Reach for Aggregated to understand an agent's structure, and Expanded to walk through a single run.
GitHub Discussions
Last edited
