OPEN LINK ↗
// 42m agoRESEARCH PAPER
EnvOpt Cuts Agent Costs, Boosts Pass Rates
EnvOpt introduces automated optimization of stationary agent environments, restructuring information and tools without changing the target model or harness. The paper reports a 72.7% cost reduction and 14.34% pass-rate gain for Claude Haiku 4.5 on AutomationBench.
// ANALYSIS
EnvOpt makes environment design a serious optimization frontier, though its striking benchmark gains still need validation in messy, changing production workspaces.
- –Targets information layout, indexing, tool interfaces, and shared context
- –Transfers gains across Claude Haiku 4.5 and Opus 5
- –Works alongside harness optimization rather than replacing it
- –Avoids requiring graders or reference answers during optimization
- –Results are author-reported and may not generalize beyond benchmark environments
// TAGS
envoptagentcontext-engineeringevaluationbenchmarkresearchopen-source
DISCOVERED
42m ago
2026-11-08
PUBLISHED
45m ago
2026-11-08
RELEVANCE
9/ 10
AUTHOR
Discover AI