RESEARCH · RESEARCH · #3194
EnvOpt: automated optimization for stationary agent environments (arXiv:2611.07372v1)
EnvOpt is a method for automated end-to-end optimization of stationary environments for LLM agents that works with any model and harness and does not require graders or reference answers. On AutomationBench it reduced Claude Haiku 4.5's cost by 72.7% and raised its pass rate by 14.34%, cut Opus 5's cost by 68.1% with a 1.34% gain, provided an additional 83.4% cost reduction and 11.17% accuracy improvement when layered on Meta-Harness for Haiku 4.5, and yielded consistent cost reductions across five benchmarks while transferring between models and harnesses.
KEY POINTS
- EnvOpt is a method for automated end-to-end optimization of stationary environments for LLM agents that works with any model and harness and does not require graders or reference answers.
- On AutomationBench it reduced Claude Haiku 4.5's cost by 72.7% and raised its pass rate by 14.34%, cut Opus 5's cost by 68.1% with a 1.34% gain, provided an additional 83.4% cost reduction and 11.17% accuracy improvement when layered on Meta-Harness for Haiku 4.5, and yielded consistent cost reductions across five benchmarks while transferring between models and harnesses.
- Demonstrates that optimizing how information is represented to agents — the environment — is a distinct, high-impact axis (beyond model and harness tuning) that can substantially lower costs and improve success without labeled data.
WHY IT MATTERS
Demonstrates that optimizing how information is represented to agents — the environment — is a distinct, high-impact axis (beyond model and harness tuning) that can substantially lower costs and improve success without labeled data.