The cost and success of an LLM agent depend jointly on its model, its harness, and the environment it operates in. Beyond the well‑studied dimensions of parameter tuning and harness optimization, we introduce a third direction: optimizing stationary environments—shared, relatively fixed workspaces such as a company’s Slack channel, document repository, or ticket system. EnvOpt is an automated end‑to‑end method that, given any model and harness, optimizes such environments using real or synthetic tasks without requiring graders or reference answers. On AutomationBench, EnvOpt cuts the cost of Claude Haiku 4.5 by 72.7% while raising its pass rate by 14.34%. The same optimized environment transfers to Opus 5, reducing cost by 68.1% and improving pass rate by 1.34%. When combined with Meta‑Harness’s optimized harness for Haiku 4.5, EnvOpt further lowers cost by 83.4% and adds 11.17% accuracy over Meta‑Harness alone. Across five benchmarks, EnvOpt consistently reduces agent cost while maintaining or improving pass rates, and it transfers across different models and harnesses. We also discuss common successful and rejected strategies observed during automated optimization. Overall, the results highlight that how information is presented to an agent shapes both its cost and success, making environment optimization as important as parameter and harness tuning.
Review: EnvOpt provides a compelling new angle on boosting LLM agent efficiency by reshaping the information landscape, offering a practical, reproducible path to significant cost‑performance gains in real‑world deployments.