Graph View: Aggregated and Expanded modes
Nikita KabardinThe trace graph view now offers two layouts, an aggregated overview and an expanded as-it-ran DAG, on a rebuilt deterministic renderer.
The graph view turns an agent trace into a picture of what your agent did: nodes for steps, edges for how it moved between them. It now runs on a rebuilt renderer and ships with two modes, so you can read the same run two ways:
- See an agent's shape at a glance with Aggregated: steps that share a name collapse into one node with a counter (
retrieve_docs (3/3)), and loops draw as cycles. - Follow a single run step by step with Expanded: every call is its own node, and loops unroll into a DAG in execution order.
- Keep your place on large traces: a deterministic layout draws the graph the same way every time, and the viewport stays put as you pan and zoom.
A rebuilt graph, on its own renderer
The graph runs on our own renderer: a deterministic layout (built on ELK) drawn onto a stable, zoomable canvas. The graph lays out the same way every time you open a trace, with no reshuffling between renders, and the viewport stays put as you pan and zoom, so you keep your place on a big agent. The layout is built to stay readable as traces grow.
Two modes: Aggregated and Expanded
An Aggregated / Expanded toggle sits in the top-left of the graph. It re-draws the same trace two ways: the difference is whether repeated work is grouped by name or shown call-by-call. Your choice is remembered across traces.
-
Aggregated (the view you've had, now named) is a compact overview. Steps that share a name collapse into a single node with a counter, so
retrieve_docs (3/3)means it ran three times, and loops are drawn as cycles. It is the fastest way to see an agent's overall shape and complexity without every individual call on screen. -
Expanded shows the process as it ran. Every call is its own node, so three
litellm_requestcalls are three distinct nodes, and loops unroll into an acyclic graph (a DAG) in execution order. It is the more literal view, closer to the trace tree, for following a specific run or pinning down exactly where something happened.
The two answer different questions: reach for Aggregated to understand an agent's structure at a glance, and Expanded to walk through a single execution in the order it ran. Expanded is a straight DAG, so it reads easily even on agents with lots of loops.
The graph view is in beta, and we want to hear what you'd like next.