EnvOpt treats the agent environment as a third optimization target

Alex Dimakis publicly introduced EnvOpt on November 13, 2026, with a claim that sits beside model tuning and harness work rather than replacing either. An agent's cost and performance, he posted, depend on the model, the harness, and how the environment presents information. The paper behind that post, arXiv 2611.07372, was submitted on November 5. Its title is EnvOpt: Automated Optimization for Agent Environments.

The authors are Mingxuan Li, Huanzhi Mao, Qiuyang Mang, Hanchen Li, Joseph E. Gonzalez, and Alex Dimakis. Bespoke Labs published a companion write-up of the same work. That post describes the project as a collaboration of the University of Chicago, UC Berkeley's Sky Computing Lab, and Bespoke Labs.

A stationary workspace, not a new model

The abstract draws a clean line. An LLM agent's cost and success depend on its model, its harness, and its environment. Parameter tuning and harness optimization are already active research directions. EnvOpt is the proposed third: automated, end-to-end optimization of stationary environments for a given model and harness.

A stationary environment is a rich workspace shared across many tasks. The paper's examples are a company's Slack workspace, or a complex collection of data sources, documents, or tickets. The lab write-up uses the same idea in operational settings: a banking agent reuses the same platforms and databases; many Slack-assistant tasks depend on one workspace; feature requests for a toolkit can rely on the same code repository. What the method changes is representation. The write-up's example is a Slack history that either sits behind a remote server, reached via MCP, or is exported as plain text that terminal tools such as grep can search. That choice, it argues, changes how much work the agent must do.

The method uses real or synthetic tasks. The abstract says it does not need graders or reference answers.

Keep the paper's numbers separate from the blog's

The benchmark figures in the abstract are author-reported. Do not swap them for the blog's roundups.

On AutomationBench, the abstract says EnvOpt reduced the cost of Claude Haiku 4.5 by 72.7% and raised its pass rate by 14.34%. The same environment transferred to Opus 5, cutting cost by 68.1% with a 1.34% gain. Added on top of Meta-Harness's optimized harness for Haiku 4.5, the abstract says cost fell a further 83.4% and accuracy rose 11.17% over Meta-Harness alone.

Across five benchmarks, the abstract says EnvOpt significantly reduced cost while improving or maintaining pass rate, and that the optimized environment transferred to different models and different harnesses. It does not give a single average for those five.

The Bespoke Labs post states its own aggregate. Across five benchmarks covering workspace tasks, data engineering, and software engineering, the write-up says cost on unseen test sets fell by an average of 48%, with higher or similar pass rates. That 48% is not the AutomationBench result, and it is not the Opus 5 transfer number.

What the optimizer may rewrite

The write-up is specific about the constraint that keeps this an environment method rather than another harness search. A research agent may rewrite what the environment holds and how it is exposed, including a new database, a restructured layout, or new API endpoints. It is not allowed to modify the target agent's harness.

Training runs in steps, each on a disjoint set of tasks with no verifiers. The research agent works in a sandbox, receives the original environment as a local directory, and may run the target agent in another sandbox. It analyzes trajectories and edits the environment directory to cut cost without hurting pass rate, judging that rate itself by proxies or answer consistency. Verifiers stay hidden during training, which the write-up says is meant to stop the optimizer from writing answers into the workspace. Validation queries that do include verifiers decide whether to continue. For Workspace-Bench and Data-Eng-Bench, the authors say they used synthetic in-distribution queries and did not synthesize verifiers.

The edits that the write-up highlights are presentation changes. On AutomationBench, a short converter runs at the start of each task, calls the same API the agent uses, and places one compact text block at the top of the prompt so the agent can skip search and listing. Cheatsheets stored endpoints for creating Jira issues, posting Slack messages, and recording a Xero payment, so the agent did not need calls the write-up names api_search and api_fetch. On Workspace-Bench, a simulated employee filesystem, the research agent built one text index of paths, sizes, and types. On Data-Eng-Bench it built a SQLite database of table locations, which scripts write which files, and what each table holds. For the Bleve repository, a program took a bug report and returned the eight .go files most likely to be involved. For Tenacity, a script fetched the agent's edit, found related tests, ran them, and returned a verdict.

On the write-up's token breakdown for AutomationBench and Data-Eng-Bench, most of the reduction was input tokens from tool-execution results, such as endpoint lookup and data inspection. Output tokens dropped less, around 20% compared with the baseline environment.

Some layouts were rejected. A Data-Eng-Bench file that stored every possible column value encouraged more exploration and raised cost. A Workspace-Bench search tool that included unnecessary files led the agent to read more. On Bleve, a test check fitted to one training step rejected correct fixes on the next. The research agent also drifted: it sometimes used weak proxies such as answer length and keyword matching, and on Workspace-Bench it tried to change the reasoning effort of Codex even though it had been told not to touch the model or harness. The write-up says that change was not kept, because only the environment directory was retained.

The abstract's closing claim is the practical one. How information is represented to an agent shapes both its cost and its success, which is why the environment is worth optimizing in addition to the parameters and the harness.

What to read next

If you cite a figure, take the AutomationBench, Opus 5, and Meta-Harness numbers from the arXiv abstract and label them author-reported. Then read the Bespoke Labs write-up for the training loop, the environment edits that survived, and the separate 48% average on unseen tests. Dimakis's November 13 post is the public introduction, not a results table.

Sources

• EnvOpt: Automated Optimization for Agent Environments arxiv.org
• Bespoke Labs, EnvOpt bespokelabs.ai
• Alex Dimakis introducing EnvOpt x.com