EnvOpt: Your Agent's Environment Is the New Model

🔥 What happened Researchers from Berkeley and UT Austin dropped EnvOpt, a method that automatically optimizes a stationary agent environment end-to-end — no graders, no reference answers needed. On AutomationBench it cut Claude Haiku 4.5's cost by 72.7% while lifting its pass rate 14.34%. 💡 Why it matters Teams have obsessed over parameters and harnesses. EnvOpt proves the environment itself is a third, untapped lever: how information is represented to an agent shapes both cost and success. The optimized environment even transfers to Opus 5, slashing its cost another 68.1%. ⚡ Our take If you're only tuning the model, you're leaving roughly 70% of the cost on the table. Harness vendors should be sweating — the environment just became the product.