The idea behind this article came across my mind when I took the LangGraph course and while working on the first examples, it kept me busy doing some fun work with the Graph workflow. Then I realized how helpful these stateful graphs are and can deliver insight into the flow of requests and queries within a multi agent AI system.
The purpose of this article is to demonstrate a probabilistic analysis on some most common agentic workflows in particular using LangGraph. Let's consider a general example of a multi agent AI system with an LLM as the decision maker to do the following tasks: Stateflow of the Agentic AI SystemHere is the breakdown of what is happening inside this workflow:
- The user query is passed to the first node Refine Query
- Based out of the context, the LLM routes the response to a Search or Generate Summary and
- Finally the process is complete when the LLM generates the response.
This is a very simple representation of an agentic workflow. In the following, I show a high level representation of this process using LangGraph.
Note: You don't need to call ANY LLMs or APIs to run this workflow. It's all running hypothetical scenarios.
We mock the above workflow using the following scenario:
-
State Representation: Uses a typed dictionary (TypedDictState) to track core variables (e.g., tools_called, name, condition) along with performance metrics (e.g., steps, node_visits, termination_reason).
-
Node Functions: node_1: Records visits and updates the name. node_2: Increments tools_called, sets condition to "good", and records its visit. - It represents of a fact that the LLM calls a search function or another tool. node_3: Similar to node_2 but sets condition to "bad" and marks termination. - It acts like a final sentiment labeling of the generated response and sends it to the user.
-
Decision Functions: decide_condition: Routes 95% to node_2 and 5% to node_3. - node_3 is a bad labeler so chances of visiting this label is low. Better to check with node_2 because this node has a better performance i.e. labels the state as good! decide_condition_2: Terminates if tools_called > 55, else loops back to node_1. - This condition is set to terminate the operation for a particular number of iterations i.e. enforce reaching the END statement. Graph Construction:
-
Assembles nodes and transitions with LangGraph starting at START and terminating upon reaching an absorbing condition (via node_3 or exceeding tool limits). Termination:
-
Ends when a bad outcome is reached or resource usage exceeds limits, enabling analysis of performance and decision efficiency.
Here is a simple Graph workflow (no LLM or anything expensive needed).
