🤖 Build resilient AI workflows with automatic GPT and Gemini failover chain

This workflow contains community nodes that are only compatible with the self-hosted version of n8n.

How it works

This workflow demonstrates how to create a resilient AI Agent that automatically falls back to a different language model if the primary one fails. This is useful for handling API errors, rate limits, or model outages without interrupting your process.

1. State Initialization: The Agent Variables node initializes a fail_count to 0. This counter tracks how many models have been attempted.
2. Dynamic Model Selection: The Fallback Models (a LangChain Code node) acts as a router. It receives a list of all connected AI models and, based on the current fail_count, selects which one to use for this attempt (0 for the first model, 1 for the second, etc.).
3. Agent Execution: The AI Agent node attempts to run your prompt using the model selected by the router.
4. The Fallback Loop:
   • On Success: The workflow completes successfully.
   • On Error: If the AI Agent node fails, its "On Error" output is triggered. This path loops back to the Agent Variables node, which increments the fail_count by 1. The process then repeats, causing the Fallback Models router to select the next model in the list.
5. Final Failure: If all connected models are tried and fail, the workflow will stop with an error.

Set up steps

Setup time: ~3-5 minutes

1. Configure Credentials: Ensure you have the necessary credentials (e.g., for OpenAI, Google AI) configured in your n8n instance.
2. Define Your Model Chain:
   • Add the AI model nodes you want to use to the canvas (e.g., OpenAI, Google Gemini, Anthropic).
   • Connect them to the Fallback Models node.
   • Important: The order in which you connect the models determines the fallback order. The model nodes first created/connected will be tried first.
3. Set Your Prompt: Open the AI Agent node and enter the prompt you want to execute.
4. Test: Run the workflow. To test the fallback logic, you can temporarily disable the First Model node or configure it with invalid credentials to force an error.