🔍Dataiku's Kiji Inspector Opens Up Why AI Agents Pick Their Tools

Your agents pick tools for reasons you can finally see

Dataiku's 575 Lab just open-sourced Kiji Inspector, and it tackles the question everyone running AI agents keeps dodging: why did it pick that tool? Prompt engineering, behavioral testing, and chain-of-thought explanations don't give you causality. Kiji Inspector does. If you're building agents that call APIs, databases, or internal services, this is aimed squarely at you. Software architects and senior engineering leads can use it to trace which internal features drove a tool selection, then feed that into production monitoring. The catch? You need to be comfortable with the underlying interpretability approach, which is not a weekend project. The method uses sparse autoencoders. Standard autoencoders squeeze inputs into a smaller latent vector and reconstruct them. Sparse autoencoders go wide instead, forcing the model to represent information through a limited set of distinct features. That separation pulls apart polysemantic context, so one neuron no longer means five things at once. Kiji Inspector reconstructs activations to surface the themes behind each decision. Worth watching: how well this scales beyond the examples Dataiku has shown, and whether other labs adopt the same approach.