Bounded Autonomy Architecture: Typed Action Contracts Constrain LLM Execution at the System Level

arXiv·medium signalXBlueskyLinkedInCopy link

A new paper (arXiv 2604.14723) argues that unsafe LLM agent behavior is primarily an execution architecture problem, not a model alignment problem. The proposed bounded-autonomy architecture lets LLMs interpret intent and propose actions, but all executable behavior is constrained by typed action contracts with consumer-side execution — the model never directly executes anything. For enterprise agent builders, this is a concrete architectural pattern: typed schemas, execution sandboxing, and contract validation as the safety layer, rather than relying on prompt engineering or RLHF.