Request for Example: Combining Human Interaction with Agent Handoff in LangGraph4J

[chenypgg]

Problem Description
When developing complex LLM applications using LangGraph4J, we often organize business modules as sub-agents (e.g., AssetAgent, OrderAgent) using the Agent Handoff pattern. Each sub-agent exposes specific functionalities through Tools. However, some functionalities require human interaction (e.g., multi-step form filling), which necessitates launching a subgraph with interruptible nodes from within a Tool method.

The key challenge is: How to synchronize state between a human-interactive subgraph and the main agent workflow, specifically:

• When a subgraph enters an interrupted state (awaiting human input), the main agent workflow should also pause
• After user input is received, both the subgraph and main agent should resume execution
• State persistence across this handoff process

[bsorrentino]

hi @chenypgg thanks for valuable feedback

Your use case is very interesting and challenging. The subgraph interruptions is a great feature but complex to implement because implies to maintain an interruptions stack. I've implemented subgraph as state graph properly for try to implement such feature and I don't know if it is enough for your use case.

In agent-handoff approach the interruption is not the right solution because the state is not shared with the action/tool so we have only arguments/return on which working on.

If you like agent-handoff pattern, you must implement interruption through arguments and result, for example you must pass an extra argument where specify if it is a resume call and its return must contain an extra information that inform that the there was an interruption

I can think about a reference implementation about this topic

[bsorrentino]

Hi @chenypgg I've added the feature Share LangGraph4j's state to tools that allow to implements what you asked for