EnvOpt: Automated Optimization for Agent Environments

EnvOpt, a method from University of Chicago, UC Berkeley Sky Computing Lab and Bespoke Labs, automatically restructures an agent's stationary environment to cut cost, reducing target-agent cost on unseen test sets by an average of 48% across five benchmarks with higher or similar pass rates. On AutomationBench it beats Meta-Harness by 9.67% pass rate at only 15% of its cost, and the optimized environments transfer to other harnesses and models, including a 75.7% average cost reduction on AutomationBench. Most savings come from input tokens, mainly tool execution results, and an offline optimization on AutomationBench costs around USD $150 while saving 77.5% of the cost of running Claude Opus 4.8.