Li, Mingxuan, Mao, Huanzhi, Mang, Qiuyang, Li, Hanchen, Gonzalez, Joseph E., Dimakis, Alex · Nov 5, 2026
EnvOpt authors report Claude gains from optimizing agent environments
Markdown · JSON · Original source
The authors of EnvOpt argue that an agent’s environment—not only its model and harness—can be optimized to reduce cost and improve task success. They define stationary environments as shared settings, such as a company’s Slack workspace or collections of documents and tickets, where agents handle many tasks. EnvOpt automates changes to these environments for a given model and harness, using real or synthetic tasks without graders or reference answers. On AutomationBench, the authors report lower costs and higher pass rates for Haiku 4.5; they also report cost reductions for Opus 5 and additional gains when EnvOpt is combined with Meta-Harness’s optimized harness. Across five , they say EnvOpt reduces agent costs while improving or maintaining pass rate, and transfers across models and harnesses. Their central point is that how information is represented to an agent can affect both its cost and success, making the environment another optimization target. The abstract does not detail the experiments or the strategies that succeeded or were rejected.
Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.
Original source
cs.AI · Article
Source details
Original source date: Nov 5, 2026 (2026-11-05).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-08T22:16:23.999Z. First observed: 2026-11-06T05:09:55.500Z. Retrieved: 2026-11-08T22:11:14.584Z.
Referenced by
Recorded citations and public file attachments. A reference alone does not establish agreement or influence.
No public references recorded here yet.
Limits
The supplied evidence is the paper’s abstract; it does not give experimental details, -by-benchmark results, or the common successful and rejected strategies mentioned by the authors.
Open question
How do the reported cost and pass-rate results vary across the five and the different environment optimization strategies?
Source-stated
The authors propose optimizing stationary environments as a third direction alongside model parameter tuning and harness optimization.
we propose the third direction: optimization over stationary environments.
Source-stated
EnvOpt is described as optimizing environments for a given model and harness using real or synthetic tasks without graders or reference answers.
using real or synthetic tasks without needing any graders or reference answers.
Source-stated
On AutomationBench, the authors report reduced cost and increased pass rate for Claude Haiku 4.5.
On AutomationBench, EnvOpt reduces the cost of Claude Haiku 4.5 by 72.7% and raises its pass rate by 14.34%.
Source-stated
The authors report that EnvOpt reduces agent costs while improving or maintaining pass rate across five benchmarks and transfers across models and harnesses.
Across 5 benchmarks, EnvOpt significantly reduces the cost of agents while improving or maintaining pass rate, and is transferable to different models and different harnesses.
Source-stated
The authors argue that information representation in the environment can affect both agent cost and success.
how information is represented to an agent shapes both its cost and its success
Source and contributor context
Li, Mingxuan, Mao, Huanzhi, Mang, Qiuyang, Li, Hanchen, Gonzalez, Joseph E., Dimakis, Alex · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Benchmark
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Claude
Claude is Anthropic’s family of AI models and assistants for tasks such as writing, analysis and coding.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Editorial provenance
Version 1. Publication ID: c1b62f4b-409a-4cf1-8915-a64ca9bb5d4b.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: passed by anthropic/claude-opus-5.5 on 2026-11-08T22:16:23.999Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 8dcdd05039bf5b929cb3508a70b5fa6818f3a9842c4f198a561a5247b20b486c. Source text hash: cf11755910ce033daa24a87235c2f52e89a3d267ff65e4a9af50f72d018f2b9a.
Revision and source-check history
Version 1 · 2026-11-08T22:16:23.999Z
EnvOpt authors report Claude gains from optimizing agent environments
The authors of EnvOpt argue that an agent’s environment—not only its model and harness—can be optimized to reduce cost and improve task success. They define stationary environments as shared settings, such as a company’s Slack workspace or collections of documents and tickets, where agents handle many tasks. EnvOpt automates changes to these environments for a given model and harness, using real or synthetic tasks without graders or reference answers. On AutomationBench, the authors report lower costs and higher pass rates for Haiku 4.5; they also report cost reductions for Opus 5 and additional gains when EnvOpt is combined with Meta-Harness’s optimized harness. Across five , they say EnvOpt reduces agent costs while improving or maintaining pass rate, and transfers across models and harnesses. Their central point is that how information is represented to an agent can affect both its cost and success, making the environment another optimization target. The abstract does not detail the experiments or the strategies that succeeded or were rejected.
Original source
cs.AI · Article
Source details
Original source date: Nov 5, 2026 (2026-11-05).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-08T22:16:23.999Z. First observed: 2026-11-06T05:09:55.500Z. Retrieved: 2026-11-08T22:11:14.584Z.
Limits
The supplied evidence is the paper’s abstract; it does not give experimental details, -by-benchmark results, or the common successful and rejected strategies mentioned by the authors.
Open question
How do the reported cost and pass-rate results vary across the five and the different environment optimization strategies?
Source-stated
The authors propose optimizing stationary environments as a third direction alongside model parameter tuning and harness optimization.
we propose the third direction: optimization over stationary environments.
Source-stated
EnvOpt is described as optimizing environments for a given model and harness using real or synthetic tasks without graders or reference answers.
using real or synthetic tasks without needing any graders or reference answers.
Source-stated
On AutomationBench, the authors report reduced cost and increased pass rate for Claude Haiku 4.5.
On AutomationBench, EnvOpt reduces the cost of Claude Haiku 4.5 by 72.7% and raises its pass rate by 14.34%.
Source-stated
The authors report that EnvOpt reduces agent costs while improving or maintaining pass rate across five benchmarks and transfers across models and harnesses.
Across 5 benchmarks, EnvOpt significantly reduces the cost of agents while improving or maintaining pass rate, and is transferable to different models and different harnesses.
Source-stated
The authors argue that information representation in the environment can affect both agent cost and success.
how information is represented to an agent shapes both its cost and its success
Source and contributor context
Li, Mingxuan, Mao, Huanzhi, Mang, Qiuyang, Li, Hanchen, Gonzalez, Joseph E., Dimakis, Alex · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Benchmark
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Claude
Claude is Anthropic’s family of AI models and assistants for tasks such as writing, analysis and coding.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Editorial provenance
Version 1. Publication ID: c1b62f4b-409a-4cf1-8915-a64ca9bb5d4b.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: passed by anthropic/claude-opus-5.5 on 2026-11-08T22:16:23.999Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 8dcdd05039bf5b929cb3508a70b5fa6818f3a9842c4f198a561a5247b20b486c. Source text hash: cf11755910ce033daa24a87235c2f52e89a3d267ff65e4a9af50f72d018f2b9a.
passed · 2026-11-08T22:16:23.999Z · anthropic/claude-opus-5.5
Publication: c1b62f4b-409a-4cf1-8915-a64ca9bb5d4b
attribution — pass: The reading attributes the proposal, method and results to the EnvOpt authors, using phrases such as 'the authors report' and 'they say'. Benchmark figures stay framed as the authors' claims, not established facts.
Authors:Mingxuan Li, Huanzhi Mao, Qiuyang Mang, Hanchen Li, Joseph E. Gonzalez, Alex Dimakis
summary — pass: The reading leads with the authors' thesis that the environment is a third optimization target, alongside model parameters and the harness. It defines stationary environments plainly and preserves the scope of the results. 'Harness' and 'AutomationBench' are left as exact technical terms; harness is a candidate for a term-library tooltip if one is missing. The title and question follow from the abstract. The reading correctly notes that the abstract mentions, but does not detail, the successful and rejected strategies.
An LLM agent's cost and success depend on its model, its harness, and its environment.
date — pass: The displayed original publication date of 2026-11-05 matches the arXiv v1 submission date. The feed date of Nov 6 is later and is correctly not used.
[Submitted on 5 Nov 2026]
support — pass: Each claim matches a verbatim abstract excerpt. The Opus 5 cost reduction is supported: the abstract reports a cost cut of 68.1%. The Meta-Harness combined gains are also supported. The reading invents no benefits, costs or real-world reliability claims, and its limitation note is accurate.
The same environment transfers to Opus 5, cutting its cost by 68.1% with a 1.34% gain.
media — pass: The selected media is the verified direct link to the original arXiv abstract page, which is the most direct access to the source.
Challenge or correct this reading
Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.
Claude is Anthropic’s family of AI models and assistants for tasks such as writing, analysis and coding.
Explore ClaudeA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referencearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceClaude is Anthropic’s family of AI models and assistants for tasks such as writing, analysis and coding.
Explore ClaudeA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referencearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition reference