LangGraph's Fault Tolerance: Retry, Timeout, Error Handler

AI agents in real production environments face various errors that are hard to see at the prototype stage, such as network failures, tool call errors, and LLM rate limits. When an unrecoverable error occurs while a task is in progress, restarting the entire process from scratch is not a sustainable way to operate.

LangGraph offers fault tolerance as a core feature to address this problem. Since LangGraph composes each step of an agent as a node and directly manages execution control, it is optimized for handling failure situations at each step.

LangGraph provides three core primitives for fault tolerance:

• RetryPolicy: Automatically retries transient errors by applying exponential backoff and jitter.
• TimeoutPolicy: Sets limits on node execution time based on actual time or progress.
• error_handler: A node that runs with the failure context included after all retries have been exhausted.

These features can be configured directly on nodes via add_node, allowing fault tolerance settings to be managed right next to the business logic. In particular, RetryPolicy is designed to be efficient by retrying transient errors like ConnectionError or 5xx responses, while not retrying programming bugs like ValueError or TypeError.