Reorganizing an agent’s data cuts AI costs 48%, researchers say

Key Points

1. A new arXiv paper reports AI agent costs fell roughly 48% on average after optimizing the shared environment.
2. Prior work tuned models and harnesses; this targets the files, tools and data agents read.
3. Gains transferred across models and harnesses, suggesting cheaper agents without switching vendors or rebuilding stacks.

The latest:

Automatically restructuring the files, tools and data an AI agent works with cut costs by about 48% on average across five benchmarks, according to a paper posted on arXiv on November 5, 2026. On AutomationBench with Claude Haiku 4.5, the authors report cost fell 72.7% while success rose 14.34%. They argue the environment deserves optimization alongside the model and the harness.

Details:

• The method: EnvOpt optimizes what the authors call a stationary environment — shared resources such as a company Slack workspace, data sources, documents or tickets reused across many tasks. The system works end-to-end for a given model and harness, runs on real or synthetic tasks, and requires no graders or reference answers.
• How it runs: A research agent such as Claude Code or Codex executes the target agent on separate training sets without verifiers, edits the environment, and keeps only changes that lower cost while holding accuracy steady using proxy metrics or answer consistency. A validation set with verifiers decides when to stop. Harness edits are excluded by design.
• What it changes: Documented edits include rendering Gmail data as plain text, converting xlsx to txt, checklists for API endpoints and command templates, a single file index mapping paths, sizes and types, a SQLite database describing tables and their sources, a ranker for files likely tied to a bug report, and an automated test script.
• The headline numbers: On AutomationBench with Haiku 4.5, cost dropped 72.7% and success rose 14.34%. Moving the same optimized environment to Opus 5 produced a 68.1% cost reduction with a 1.34% gain. Stacked on the Haiku-tuned Meta-Harness, EnvOpt added 83.4% further cost savings and 11.17% higher accuracy.
• The other benchmarks: Workspace-Bench, built on personal workspaces, showed 47.65% lower cost and 3.3% higher success. Data-Eng-Bench, covering dbt data engineering, saw 53.44% lower cost with no performance loss. On the Bleve repository the agent reached near-perfect accuracy in 56.9% of gskill’s completion time.
• Tool ablation: On Tenacity repository tasks drawn from Advisor Models, EnvOpt reported 8.9% higher success with fewer steps. An ablation found that adding new tools or endpoints alone helped, but less than tools combined with broader environment changes — the authors’ evidence that formatting, not just capability, drives the savings.
• What failed: Packing too much into one file increased exploration on Data-Eng-Bench. A search tool returning too many files raised reading cost on Workspace-Bench. A test check fitting only some training tasks rejected correct fixes on Bleve. An early fault in a SQL splitter went unnoticed for four steps on Data-Eng-Bench.
• Who wrote it: The authors are Mingxuan Li, Huanzhi Mao, Qiuyang Mang, Hanchen Li, Joseph E. Gonzalez and Alex Dimakis, describing the project as a collaboration between the University of Chicago, Berkeley’s Sky Computing Lab and Bespoke Labs. The abstract does not specify whether success gains are percentage points or relative percentages, and no code release link appears.

Background:

Work on cutting agent costs has concentrated on two levers: tuning model parameters and improving the harness, the scaffolding that orchestrates an agent’s steps. The paper positions the environment — the data and tools the agent actually reads — as a third, largely unexploited lever.

Between the lines:

The transfer results carry the commercial weight. An environment optimized for Haiku 4.5 still delivered 68.1% savings on Opus 5, and stacking on an already-tuned Meta-Harness added 83.4% more — meaning the work survives a model upgrade rather than being discarded with it. The documented failures, from the over-stuffed file to the unreliable test check, mark where automated optimization still needs a human check.

What’s next

Watch for a code release, independent replication of the 48% average across other agent frameworks, and whether the success-rate figures are clarified as percentage points or relative gains in a revised version.