Semantic Early-Stopping Replaces the Fixed Iteration Cap in LLM Agent Loops

arXiv 2606.27009·high signal

This paper targets the near-universal max_iterations kill-switch in Writer/Critic-style agent loops, calling it a 'syntactic' stop that over-spends tokens on easy inputs and truncates hard ones. The proposed alternative halts when consecutive draft embeddings stop changing in meaning (cosine distance with a patience window) plus a quality check. It's directly actionable for builders running iterative refinement loops who want cost control tied to convergence rather than an arbitrary count.