EnvOpt
EnvOpt is a method for automatically optimizing an agent's static environment. It requires neither a scorer nor reference answers, relying solely on real or synthetic tasks to optimize the environment representation end-to-end, thereby reducing agent cost and improving success rates, with optimization results transferable across models and harnesses.
Timeline (last 90 days)
EnvOpt is a method that optimizes an agent's static environment end to end using only real or synthetic tasks, without the need for labeled answers or graders.
EnvOpt does not change model weights or rewrite the harness; instead, it adjusts the information world the agent faces.
On AutomationBench, EnvOpt reduced the cost of Claude Haiku 4.5 by 72.7% while increasing the pass rate by 14.34%.
When the same optimized environment is used directly with Opus 5, cost drops by 68.1% while the pass rate still grows by 1.34%.
When EnvOpt is layered on top of the harness that Meta-Harness optimized for Haiku 4.5, cost drops by a further 83.4% and accuracy improves by an additional 11.17% compared with using Meta-Harness alone.
Across 5 benchmarks, EnvOpt significantly reduces agent cost while maintaining or improving the pass rate, and it can transfer to different models and different harnesses.
EnvOpt adapts to any given model and harness.
The EnvOpt paper discusses successful strategies and rejected strategies commonly seen in the automated optimization process.
All Facts (8)
EnvOpt is a method that optimizes an agent's static environment end to end using only real or synthetic tasks, without the need for labeled answers or graders.
50%UnverifiedEnvOpt does not change model weights or rewrite the harness; instead, it adjusts the information world the agent faces.
50%UnverifiedOn AutomationBench, EnvOpt reduced the cost of Claude Haiku 4.5 by 72.7% while increasing the pass rate by 14.34%.
50%UnverifiedWhen the same optimized environment is used directly with Opus 5, cost drops by 68.1% while the pass rate still grows by 1.34%.
50%UnverifiedWhen EnvOpt is layered on top of the harness that Meta-Harness optimized for Haiku 4.5, cost drops by a further 83.4% and accuracy improves by an additional 11.17% compared with using Meta-Harness alone.
50%UnverifiedAcross 5 benchmarks, EnvOpt significantly reduces agent cost while maintaining or improving the pass rate, and it can transfer to different models and different harnesses.
50%UnverifiedEnvOpt adapts to any given model and harness.
50%UnverifiedThe EnvOpt paper discusses successful strategies and rejected strategies commonly seen in the automated optimization process.
50%