Agent Execution Graph
A step-through, clickable trace of the ReAct loop — Thought → Action → Observation, repeating until the model has enough to answer. Unlike the other visualizations in this lab, this one is a scripted, representative trace, not a live model call — there's no LLM running in your browser here. The point is to make the shape of the loop something you step through and click, not something you only read a diagram of.
Interactive
Agent Execution Graph
Scenario
“What’s the current weather in the capital of France, in Fahrenheit?”
Step 1 / 8 -- Thought
I need the capital of France, then its current weather, then I’ll need to convert Celsius to Fahrenheit.
The model reasons about what it knows so far and what it needs to do next -- no tool call, just internal reasoning text.
Trace so far
Thought: I need the capital of France, then its current weather, then…
A scripted, representative ReAct trace -- not a live model call -- so the shape of the reason / act / observe loop is something you can step through and click, not just read about.
What to Try
- Hit Step repeatedly and watch the active node move around the loop — Thought → Action → Observation → back to Thought — until the model exits the loop into Answer instead of looping again.
- Switch scenarios: the "Company lookup" example needs two full loop iterations (first find the company, then find its employee count) before it has enough to answer — a concrete example of why multi-hop questions need more than one Action.
- Click any entry in the trace log to jump back to it and re-read that exact step's detail.
- Hit Play to watch the full loop run on its own, then compare it to the Plan-and-Execute pattern described on the same page — Plan-and-Execute would commit to all the steps up front instead of deciding one Action at a time based on what the last Observation actually returned.
Back to Visual Lab Overview.