Graph State Manipulation for Multi-Agent Systems and Tools

[makoit]

I'm implementing an agent system and facing challenges in effectively manipulating the graph state within my tools. Here are the scenarios:

1. Multi-Agent System with Supervisor:

   • I have a system where a supervisor agent decides which agent to activate next. Each agent has its own tool. For example:
     • Agent A captures an image using a camera tool.
     • Agent B analyzes the image content using an analysis tool.
   • I need a way to share information (e.g., a Pillow image) between tools through the graph state, allowing one agent with a tool to write data and another agent with a tool to read it.

2. React Agent with Multiple Tools:

   • In this scenario, a single agent is responsible for using multiple tools, and I want each tool to be capable of manipulating the graph state similarly to scenario 1.

Request: What is the recommended approach for enabling agents and tools to read/write to the graph state in these scenarios?

I tried to follow this approach: pass-run-time-values-to-tools. But this does not work when I try to change the state in the tool in a way that a other tool can read the state again.

[makoit]

After implementing the whole thing by my own I recognized that finally there is a official support for updating graph state from a tool node: https://langchain-ai.github.io/langgraph/how-tos/update-state-from-tools/