EnvOpt treats the agent's environment as a third optimization target — 48% average cost cut across five benchmarks, and the artifacts are inspectable
A research agent rewrites the files, data, and tools an LLM agent reads — cutting token cost without touching the model or harness. All numbers are vendor-reported.

EnvOpt treats the agent's "stationary environment" — the files, data, and tools it reads — as a third optimization target alongside model weights and the harness. The framework, from Bespoke Labs with the University of Chicago and UC Berkeley's Sky Computing Lab, reports an average 48% cost reduction across five benchmarks on unseen tests. A research agent iteratively rewrites the environment to cut the target agent's token spend, and the resulting artifacts are published and inspectable.
What shipped
EnvOpt is described in arXiv:2611.07372 (submitted Nov 5, 2026) by Mingxuan Li, Huanzhi Mao, Qiuyang Mang, Hanchen Li, Joseph E. Gonzalez, and Alex Dimakis. The premise: an agent's cost and success depend on three things — model parameters, the harness, and the structure of the environment it interacts with — and only the first two are normally optimized. The same information can reach an agent in very different forms: a company's Slack history can sit behind a remote server reachable via MCP, or be exported as plain text files greppable with terminal tools. That choice alone changes how much work the agent does.
"Stationary environments" are the ones shared across many tasks: an enterprise's databases, a Slack workspace, a code repository. EnvOpt optimizes how those are organized and represented to a target agent, without touching the model or the harness.
The mechanism is a loop. A sandboxed "research agent" gets the original environment as a local directory and permission to run the target agent in a separate sandbox. It runs the target agent, analyzes the trajectories, and edits the environment directory to reduce cost without hurting pass rate.
Each optimization step uses a disjoint set of training tasks, and the verifiers for those tasks are hidden from the research agent — explicitly to stop it from reward-hacking by writing answers into files. A validation set with verifiers decides when the loop has converged.

What the research agent actually built is the most concrete part of the paper. On AutomationBench it wrote a converter that runs at the start of every task, calls the same API the agent uses to read task data, and rewrites the results into one compact text block at the top of the prompt so the agent skips search and listing steps. It also produced a cheatsheet of all the API endpoints needed to create Jira issues, post Slack messages, and record Xero payments, so the agent stops calling api_search/api_fetch.
On Workspace-Bench it built a single-file index of paths, sizes, and types across a simulated employee's filesystem. On Data-Eng-Bench it built an SQLite database mapping which tables hold what and which scripts write to which files. On the Bleve repository it wrote a program that takes a bug report and ranks the eight .go files most likely to be affected; on Tenacity it wrote a validate_tenacity.py that fetches the agent's code change, runs the relevant tests, and returns verdicts with no extra token cost.
Why it matters
The headline numbers are all vendor-reported, but they are specific. On AutomationBench, EnvOpt cuts Claude Haiku 4.5's cost by 72.7% while raising pass rate by 14.34%. The same optimized environment transfers to Opus 5, cutting cost 68.1% with a 1.34% pass-rate gain. Stacked on Meta-Harness's optimized harness for Haiku 4.5, it lowers cost a further 83.4% and adds 11.17% accuracy over Meta-Harness alone. The team reports the AutomationBench optimization itself cost about $150 and saved 77.5% of Claude Opus 4.8's execution cost — "easily saving many times the cost of running EnvOpt."
Transferability is the part that changes how you'd build. The optimized environments carry across models and harnesses: an average 75.7% cost reduction on AutomationBench across harnesses and models (Haiku 5.5 with EnvOpt beats Opus 4.8's pass rate and approaches Opus 5), 56.7% on Workspace-Bench across Codex/GPT-5.6-Sol and Claude Code/Opus 5, and 53.54% on Data-Eng-Bench on non-target agents. The environment is a lever you can pull once and reuse, rather than a per-model tuning artifact.
The paper is also unusually honest about what fails. It documents rejected optimizations: a single file cramming every database value encouraged the agent to explore more, raising cost; a search tool that returned too many files motivated extra reading; a test check added on Bleve rejected correct fixes in later steps. It also admits the research agent sometimes used unreliable proxies (answer length, keyword matching) to judge quality when no verifier was available, which hurt held-out performance until the instructions were tightened.

Caveats
Every performance number here is self-reported by Bespoke Labs and the co-authors; this is a non-peer-reviewed preprint with no independent replication. Treat the 48% average and the per-benchmark figures as claims, not established results.
The release is also partial. The GitHub repo contains the full optimization pipeline code for Workspace-Bench, plus the optimized environment artifacts for AutomationBench, Bleve, and Tenacity — you can watch the AutomationBench environment layer grow from 13 KB to 65 KB across six steps. But Data-Eng-Bench code and environments are missing ("stored on another machine that is currently down for maintenance"), and the Workspace-Bench environments live only as a 148 GB Hugging Face dataset with an empty README and a broken dataset viewer. There is no LICENSE file in the repo, so the code is all-rights-reserved by default.
One naming note: an unrelated paper called "EnvOpt: Closed-Loop Environment Optimization for Agent Post-Training" exists on OpenReview. Don't confuse the two.