Error Handling and Retries in LangGraph

Our assistant can now call tools in sequence or in parallel, weaving answers together like a skilled host managing a lively dinner party. But we have quietly assumed up to this point that every tool always works perfectly.

In the real world, this is rarely the case. APIs fail. Tools throw exceptions. Sometimes the guest you invited to the party forgets their notes or spills soup before answering the question. The host needs to handle these hiccups gracefully and not let the whole evening collapse. This is where error handling and retries come into play.

For example, imagine asking for the weather in Paris and London. The Paris forecaster answers promptly, but the London forecaster trips up. Perhaps their microphone breaks, or they just shrug. The host should not stop the conversation altogether. Instead, they might ask the London forecaster to try again. If the forecaster still cannot answer, the host should at least summarize what Paris said so the guest leaves with something useful.

This is what error handling looks like in an assistant: recover when things go wrong, retry when appropriate, and give partial but graceful answers if some tools fail.

How to implement error handling with tools

To illustrate resilience, we need something that occasionally breaks. Let’s adjust ...