EnvOpt: Automated Optimization for Agent Environments

Bespoke Labs introduced EnvOpt, an automated system that restructures stationary software environments to lower the operational costs of large language model agents without degrading performance. As detailed in bespokelabs.ai's report, the framework treats enterprise environments as active optimization targets rather than immutable backdrops, reconfiguring filesystems, API endpoints, and internal databases to minimize token overhead. Across five benchmark suites, the technique reduced agent token expenses by an average of 48 percent while sustaining or increasing benchmark pass rates.

This technical report details how EnvOpt restructures persistent operational workspaces, reviews the specialized artifacts generated by its optimizing agent, and examines empirical benchmark results across data engineering, software maintenance, and corporate platforms. Autonomous agent developers and platform architects can apply these findings to cut deployment expenditures while improving multi-step task execution.

01Stationary environments and agent work allocations

An autonomous agent's operational expenses and task success depend on model parameters, execution harnesses, and the structure of the surrounding environment. While developers actively tune model weights and harness scaffolding, the underlying operating environment is typically treated as fixed. Restructuring how information is exposed directly alters the computational effort an agent must exert to resolve complex tasks.

Stationary environments contain large collections of artifacts that remain constant across repeated tasks, such as enterprise banking databases, corporate Slack workspaces, and established code repositories. Serving data through remote server interfaces requires far more agent work than querying local text exports with simple terminal tools like grep. EnvOpt targets and refines these persistent structures.

Rather than altering the target agent's fixed harness or internal parameters, the framework optimizes environment contents and external access points. Modifications include restructuring existing tables, creating dedicated databases, and exposing concise API endpoints. Treating data presentation as configurable allows systems to eliminate redundant exploratory work without constraining the agent's baseline capabilities.

02Sandbox research agents and hidden verifiers

EnvOpt optimizes environments across multiple sequential stages using an autonomous research agent running inside an isolated sandbox filesystem. This research agent receives access to the original environment directory alongside permissions and compute budget to execute target agents within separate sandboxes. By evaluating runtime trajectories, the researcher iteratively reorganizes the environment to lower token expenses.

To eliminate reward hacking, the training pipeline hides all evaluation verifiers from the research agent during environment optimization passes. Withholding verifiers prevents the research agent from hardcoding explicit answers directly into local files to artificially minimize costs without generating real improvements. The agent instead evaluates target agent performance using consistency metrics and internal proxy measures.

Concealing verification rubrics also resolves task scarcity across complex enterprise environments by enabling the use of synthesized in-distribution training queries. Engineering teams can generate synthetic training problems without authoring automated verifiers, as demonstrated on Workspace-Bench and Data-Eng-Bench. At each step's conclusion, validation queries with automated verifiers evaluate whether the reorganized environment has converged.

03Automated cheatsheets, indexing maps, and local databases

During optimization on AutomationBench, the research agent generated automated environment converters and consolidated endpoint cheatsheets to streamline target agent execution. The converter program runs before each task, calling standard APIs to condense operational data into a concise text block placed directly in the agent's prompt. Dedicated cheatsheets catalog necessary API endpoints for Jira, Slack, and Xero.

For corporate simulations on Workspace-Bench, the research agent compiled a single plaintext index detailing file paths, dimensions, and types across the entire filesystem. The target agent queries this compact index rather than spending tokens exploring extensive directories. On Data-Eng-Bench, the research agent constructed an SQLite database tracking table schemas, storage locations, and file write dependencies.

In software engineering benchmarks, the system generated specialized helper scripts tailored to specific codebase architectures. For the Bleve repository, a file-ranking utility analyzes incoming bug descriptions and identifies the eight Go source files most likely affected. For the Tenacity repository, a custom Python script runs candidate test files and validates proposed patches.

04Input token savings and benchmark cost reductions

Evaluations across five benchmarks spanning workplace workflows, data engineering, and software engineering show EnvOpt reducing target agent costs by an average of 48 percent on unseen test sets. On AutomationBench, EnvOpt exceeded Meta-Harness pass rates by 9.67 percent while requiring only 15 percent of Meta-Harness's operating cost. Pass rates remained equal or superior across tests.

Token accounting indicates that input token savings within tool execution feedback deliver the largest overall cost reductions. Target agents expend dramatically fewer tokens processing verbose environment responses during endpoint lookups on AutomationBench and data inspections on Data-Eng-Bench. In contrast, output token usage declined by approximately 20 percent compared to baseline environments, primarily through reduced tool calls.

Offline optimization of the stationary environment on AutomationBench cost roughly 150 dollars while reducing runtime expenditures for Claude Opus 4.8 by 77.5 percent. On the Bleve software engineering repository, agents maintained near-perfect evaluation scores while utilizing only 56.9 percent of completion tokens. These savings rapidly recover the initial offline optimization investment across repeated runs.

05Cross-model transfers and combined harness gains

Environments optimized through EnvOpt transfer effectively across differing agent harnesses and underlying model architectures without re-running the optimization routine. On AutomationBench, transfers yielded average cost reductions of 75.7 percent, allowing Haiku 5.5 to exceed the pass rate of Opus 4.8 and approach Opus 5. Transfer evaluations on Data-Eng-Bench demonstrated average cost savings of 53.54 percent.

On Workspace-Bench, transfers to OpenAI Codex using GPT-5.6-Sol and Claude Code using Opus 5 achieved average cost reductions of 56.7 percent with steady pass rates. Combining EnvOpt with Meta-Harness expanded the Pareto frontier for Haiku 4.5. On Opus 5, pairing EnvOpt with Meta-Harness improved pass rates by 16 percent with substantially lower costs.

An ablation study isolating tool-only optimizations confirmed that adding new tool interfaces improves token efficiency and completion rates. However, tool enhancements alone achieved significantly smaller gains than jointly optimizing the environment and its access points. Full environment optimization captures substantial efficiency headroom that harness engineering and tool refinement cannot deliver on their own.