EnvOpt: Automated Optimization for Agent Environments

Researchers introduced EnvOpt, an automated method that optimizes stationary environments and representations for LLM agents to reduce operational costs without sacrificing pass rates. Across five benchmarks, EnvOpt reduced target agent costs on unseen test sets by an average of 48% while maintaining or improving pass rates. The system uses a research agent in a sandbox environment, guided by training tasks without verifiers, to restructure artifacts, build indexes, and create API endpoint cheatsheets. Open article →