Mechanistic Interpretability of Multi-Step Agentic Tool Use

Abstract & Key Findings

When an autonomous agent executes fifty sequential terminal commands, how does it represent its high-level objective versus immediate syntax errors? Using sparse autoencoders (SAEs) scaled to 32 million latent features on Alchemax-o3, we isolated specific circuit families responsible for self-correction and uncertainty estimation.

By monitoring these latent features in real time during inference, our Aegis runtime can detect potential goal drift or hallucinated tool arguments before a command is ever executed.