Strategic Verification for Long-Running LLM Agents

Large language model (LLM) agents are evolving from single-turn text generators into long-running systems capable of multi-step reasoning, tool use, persistent memory, and reusable skill development. Reinforcement learning from verifiable rewards (RLVR) can improve agent reliability, but existing approaches commonly treat verification as a fixed component of the reward structure, creating challenges related to sparse credit assignment, verifier reliability, reward hacking, path compliance, and verification cost. We introduce Budgeted Adaptive Verification with Asymmetric Rewards (BAVAR), a framework that formulates verification as a sequential, resource-constrained decision problem. BAVAR selectively determines what to verify, when verification is valuable, which verifier to invoke, and how verified evidence should influence learning based on uncertainty, action criticality, verifier reliability, expected verification value, and remaining computational budget. It combines reliability-gated positive process rewards with persistent penalties for path violations and extends verification to persistent memories and reusable skills. In an illustrative evaluation under matched verification budgets, BAVAR achieved 72.6% safe verified success, compared with 67.1% for uniform dense verification and 58.4% for outcome-only RLVR. Relative to dense verification, BAVAR reduced verification cost per safe success by 45.5% while using 47.8% fewer verifier tokens. It also reduced verified path violations to 7.2 per 100 trajectories, compared with 20.4 for outcome-only RLVR, and reduced successful verifier exploitation to 4.8%, compared with 15.9% for a single-judge baseline and 18.6% for outcome-only RLVR. For persistent artifacts, BAVAR achieved 88.7% held-out reuse success, compared with 81.3% for dense verification. These simulated findings suggest that adaptive allocation of verification resources can improve safe agent performance while reducing computational and verification overhead, providing a principled foundation for scalable verification of long-running autonomous agents.