Harrison Chase with LangChain released a walkthrough of an AI app that handles email triage. For those who have already gotten their hands dirty with the Lang ecosystem, the structure of the graph is most interesting.

On a high level, there's a succinct handling of tool calling. First, A message is drafted in response to an email. From there, a tool may be invoked (find meeting time, mark as read, etc.). Then, the graph can traverse to the appropriate tool.

draft_message seems to be the heavy lifter. Tool calls often return to the draft_message node. Not unlike other software design, a parent component is ultimately responsible for multiple iterations and linking between child components.

A few other observations:

1. The graph entry point is through a triage node. Their example uses an LLM to determine the next steps based on message context. This can be error-prone, but is likely mitigated by the fact that this app uses human-in-the-loop.
2. There’s a pattern for recovering from bad calls from the LLM. In this case, there is a bad_tool_call node that is responsible for rerouting to the draft_response node in the event that the agent hallucinates a tool. Another point of recovery!
3. Rewrite Node: Message generation follows two different passes to an LLM. One to write an initial draft (“What do I need to respond with?"), and then a rewrite node. (“What tone do I need to respond with?“) A useful pattern for message refinement.

You can find the graph code here. And here is the walkthrough, starting at the explanation of the graph structure.