Watch your coding agent with deterministic rules, not another LLM — LivePlan adds up to 15.2% resolve rate for $0.08 a run

arXiv 2608.06701·high signalXBlueskyLinkedInCopy link

LivePlan monitors a programming agent's trajectory with a rule-based detector that needs no model call, and only wakes an advisor LLM once it sees drift, repeated failed actions, or an imminent no-patch exit. On SWE-agent across five different LLMs it raised resolution rates by up to 15.2% (9.9% average) at roughly $0.08 extra per instance, with gains concentrated on medium and hard instances and almost no regression on runs that already worked. The transferable skill is the two-tier split: cheap deterministic monitors decide *when* to intervene, an expensive model decides *what* to say — inverting the usual 'LLM judge watches everything' pattern that eats budget and hallucinates corrections.