Real-world web agents do not operate in an unlimited environment. Their interactions are bounded by request budgets, time delays, anti-crawling restrictions, and the possibility that operations will fail. While many agentic systems are evaluated primarily on task success rates, this narrow focus can miss the costs that actually determine viability in production.
The arXiv paper proposes an agentic reinforcement learning decision model explicitly designed under multi-cost and failure-risk constraints. Instead of treating success as the sole objective, the model factors in the expense of requests, the cost of delays, the limitations imposed by anti-crawling defenses, and the chance of failure when choosing actions.
This approach underscores a broader point: evaluating web agents only by success rate gives an incomplete picture. Real deployment requires balancing effectiveness against resource limits and risk, and models that ignore these trade-offs may not translate well from benchmark to practice.