For the LangGraph Python library, we've added:

• Node local state, which lets you attach a state schema different from the graph's schema. This enables new state management patterns, including:

  • State keys private to a single node (to retain data for the next time that node is called)

  • State keys private to a few nodes (used to share information between nodes, without writing them to the overall graph state)

• Customization of the graph's input and output schemas, which help clarify which keys should be accepted as input and which keys should be returned as output when the graph runs

• Customized retry behavior per node. Retries can be configured with max attempts, exponential backoff parameters (where we default to sensible values), and criteria for retrying exceptions.

• Bug fix: Resolved an issue where async TimeoutError in a node was incorrectly raised as an async CancelledError in the graph.