๐Ÿ’ป EnvOpt: Optimizing LLM Agent Environments

Bespoke Labs, the University of Chicago, and UC Berkeley Sky Computing Lab have released a framework that automatically improves an agent's "stationary environment"โ€”files, data, and tools. Model parameters and the harness remain untouched, and across five benchmarks, the average savings on unseen tests is 48% while maintaining the pass rate.

๐ŸŒ The environment becomes a third optimization target alongside the model and the harness: the agent receives file indexes, SQLite table maps, and API hints in advance instead of searching from scratch. The effect transfers between models and harnesses.

๐Ÿ‘ค Agent costs can be reduced by changing not the model, but the delivery of data and tools: on AutomationBench, the optimization cost $150 and saved 77.5% of the execution costs for Claude Opus 4.8. The code is partially open on GitHub (itea1001/EnvOpt), and the environments are available on Hugging Face.

Source 1: https://bespokelabs.ai/blog/envopt Source 2: https://arxiv.org/abs/2611.07372