onBeforeModelInvocation

The onBeforeModelInvocation event fires immediately before the underlying LLM client makes a network request.

Usage

This is your final opportunity to modify the payload sent to the LLM.

It is incredibly powerful for:

• Changing the model or temperature dynamically based on the current agent state or recent conversation context.
• Injecting a hardcoded injectResponse (bypassing the network call entirely).
• Injecting a system injectMessage at the very end of the history array without saving it to the permanent state.

Payload Details

agent.onBeforeModelInvocation(({ config, context }) => {
    // If we've been running too long, switch to a cheaper/faster model
    if (context.state.stepCount > 5) {
        config.model = "gpt-4o-mini";
        config.temperature = 0.1;
    }
    
    // You MUST return the config object
    return config;
});