Presentation: Opening the Black Box: Critical Explainability for AI Agent Tool Selection with Kiji Inspector

Dataiku has released the open-source Kiji Inspector, a tool designed to interpret the internal decision-making processes of AI agents. By utilizing Sparse Autoencoders on model decision tokens, the software provides a technical analysis of why an agent selects specific tools. This approach aims to move beyond simple text-based explanations by using causal ablation and layer activation extraction to reveal the underlying logic within the model.