Park, Jongseok, Yu, Jiaxiang, Zhou, Eric, Hong, Tony, Kim, Sunga, Afifi, Yusuf, Wong, Ian, Stoica, Ion, Cheung, Alvin · Nov 9, 2026
Jongseok Park and coauthors report cost-efficient agent harness optimizations
Markdown · JSON · Original source
Jongseok Park and coauthors argue that an ’s harness—the interface between an and a computing environment—may matter more for execution cost than for task accuracy. They say prior studies found task-execution accuracy depends mostly on the LLM, while differences in system prompts or tools have minimal impact; those harness-side differences can still change cost substantially. The paper analyzes agent traces to find ways to reduce token use and LLM calls without changing the model or affecting task accuracy. The authors introduce Agency, a platform for implementing these optimizations across existing harnesses. In proof-of-concept tests on real-world , they report cost reductions of up to 35% without compromising accuracy. This suggests the system around a model may offer a way to lower agent-execution costs while retaining measured task accuracy. The abstract does not identify the individual optimization techniques or give benchmark-level results, so the reported maximum does not show how broadly the savings apply.
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 9, 2026 (2026-11-09).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-12T10:31:22.469Z. First observed: 2026-11-10T08:39:42.296Z. Retrieved: 2026-11-12T10:03:09.918Z.
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 evidence is the abstract; it does not describe the individual optimization techniques or provide -level results. The abstract reports maximum cost reductions but does not specify how often those reductions occurred or provide further detail on the tasks and benchmarks.
Open question
Which optimization techniques and results account for the reported cost reductions, and how consistently do they preserve task accuracy?
Source-stated
The authors say task-execution accuracy depends mostly on the LLM, while harness-side differences such as system prompts or tools have minimal impact on accuracy but can substantially affect cost.
AI agents' task-execution accuracy depends mostly on the LLM used, while harness-side variances such as system prompts or tools have minimal impact.
Source-stated
The authors report cost reductions of up to 2.6 times from harness-side variation on the same model and task, based on recent studies.
achieving cost reductions up to 2.6 times on the same model and task.
Source-stated
The authors analyze agent traces for ways to reduce token use and LLM calls without modifying the model or affecting task accuracy.
identify system opportunities to reduce token use and LLM calls without modifying the model or affecting task accuracy.
Source-stated
The authors introduce Agency as a platform for implementing harness-side optimizations on existing harnesses, and report proof-of-concept cost reductions of up to 35% across real-world benchmarks without compromising accuracy.
Using Agency, we implement proof-of-concept optimization techniques based on our analysis and achieve cost reductions up to 35% across various real-world benchmarks on existing harnesses without compromising accuracy.
Source and contributor context
Park, Jongseok, Yu, Jiaxiang, Zhou, Eric, Hong, Tony, Kim, Sunga, Afifi, Yusuf, Wong, Ian, Stoica, Ion, Cheung, Alvin · 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.
Agent harness
An agent harness is the system that lets an AI model act as an agent: it processes inputs, orchestrates tool calls and returns results. An evaluation harness runs evals end to end, including grading outputs.
AI agent
An AI system that chooses and carries out actions toward a goal, often using software tools.
Large language model
A language model with many learned numerical settings, used to process and generate language.
Benchmark
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
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: 1c7c6291-283a-4c63-818d-03c1414ca5ad.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: passed by anthropic/claude-opus-5.5 on 2026-11-12T10:31:22.469Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 99fb4544c9aeef6ca272b04b4fa17573c47b3d790a9b5b2782c40dea2ed10efd. Source text hash: 72d1d862466219e695b0669e607651e650864abe93752645868d02c5c9d09ccb.
Revision and source-check history
Version 1 · 2026-11-12T10:31:22.469Z
Jongseok Park and coauthors report cost-efficient agent harness optimizations
Jongseok Park and coauthors argue that an ’s harness—the interface between an and a computing environment—may matter more for execution cost than for task accuracy. They say prior studies found task-execution accuracy depends mostly on the LLM, while differences in system prompts or tools have minimal impact; those harness-side differences can still change cost substantially. The paper analyzes agent traces to find ways to reduce token use and LLM calls without changing the model or affecting task accuracy. The authors introduce Agency, a platform for implementing these optimizations across existing harnesses. In proof-of-concept tests on real-world , they report cost reductions of up to 35% without compromising accuracy. This suggests the system around a model may offer a way to lower agent-execution costs while retaining measured task accuracy. The abstract does not identify the individual optimization techniques or give benchmark-level results, so the reported maximum does not show how broadly the savings apply.
Original source
cs.AI · Article
Source details
Original source date: Nov 9, 2026 (2026-11-09).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-12T10:31:22.469Z. First observed: 2026-11-10T08:39:42.296Z. Retrieved: 2026-11-12T10:03:09.918Z.
Limits
The evidence is the abstract; it does not describe the individual optimization techniques or provide -level results. The abstract reports maximum cost reductions but does not specify how often those reductions occurred or provide further detail on the tasks and benchmarks.
Open question
Which optimization techniques and results account for the reported cost reductions, and how consistently do they preserve task accuracy?
Source-stated
The authors say task-execution accuracy depends mostly on the LLM, while harness-side differences such as system prompts or tools have minimal impact on accuracy but can substantially affect cost.
AI agents' task-execution accuracy depends mostly on the LLM used, while harness-side variances such as system prompts or tools have minimal impact.
Source-stated
The authors report cost reductions of up to 2.6 times from harness-side variation on the same model and task, based on recent studies.
achieving cost reductions up to 2.6 times on the same model and task.
Source-stated
The authors analyze agent traces for ways to reduce token use and LLM calls without modifying the model or affecting task accuracy.
identify system opportunities to reduce token use and LLM calls without modifying the model or affecting task accuracy.
Source-stated
The authors introduce Agency as a platform for implementing harness-side optimizations on existing harnesses, and report proof-of-concept cost reductions of up to 35% across real-world benchmarks without compromising accuracy.
Using Agency, we implement proof-of-concept optimization techniques based on our analysis and achieve cost reductions up to 35% across various real-world benchmarks on existing harnesses without compromising accuracy.
Source and contributor context
Park, Jongseok, Yu, Jiaxiang, Zhou, Eric, Hong, Tony, Kim, Sunga, Afifi, Yusuf, Wong, Ian, Stoica, Ion, Cheung, Alvin · 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.
Agent harness
An agent harness is the system that lets an AI model act as an agent: it processes inputs, orchestrates tool calls and returns results. An evaluation harness runs evals end to end, including grading outputs.
AI agent
An AI system that chooses and carries out actions toward a goal, often using software tools.
Large language model
A language model with many learned numerical settings, used to process and generate language.
Benchmark
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
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: 1c7c6291-283a-4c63-818d-03c1414ca5ad.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: passed by anthropic/claude-opus-5.5 on 2026-11-12T10:31:22.469Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 99fb4544c9aeef6ca272b04b4fa17573c47b3d790a9b5b2782c40dea2ed10efd. Source text hash: 72d1d862466219e695b0669e607651e650864abe93752645868d02c5c9d09ccb.
passed · 2026-11-12T10:31:22.469Z · anthropic/claude-opus-5.5
Publication: 1c7c6291-283a-4c63-818d-03c1414ca5ad
attribution — pass: The reading attributes the argument and results to Jongseok Park and coauthors, who are listed as the paper's authors. It keeps the claims about prior findings framed as what the authors say recent studies show, and it frames reported results as the authors' reports rather than established fact.
Authors:Jongseok Park, Jiaxiang Yu, Eric Zhou, Tony Hong, Sunga Kim, Yusuf Afifi, Ian Wong, Ion Stoica, Alvin Cheung
summary — pass: The summary leads with the authors' thesis and explains 'harness' in plain language. It then covers the trace analysis, the Agency platform and the up-to-35% result, and closes with a hedged significance statement and specific unknowns. The question follows from a real gap in the abstract. A minor optional repair is to name the source's examples, Claude Code and OpenAI Codex, so lay readers can picture which harnesses are meant. 'Token' and 'LLM' are standard terms suitable for tooltips.
existing off-the-shelf harnesses such as Claude Code or OpenAI Codex
date — pass: The originalPublished date of 2026-11-09 matches the arXiv v1 submission date shown on the page. It predates the feed entry date of 2026-11-10.
[Submitted on 9 Nov 2026]
support — pass: Each claim matches the abstract. The 2.6x figure is reasonably linked to the cited recent studies. The 35% result keeps its 'up to' and proof-of-concept qualifiers. The concluding 'may offer a way' sentence is conditional and does not overreach. The limitations correctly note that the abstract does not give per-technique or per-benchmark detail. No unsupported benefits or costs are introduced.
achieve cost reductions up to 35% across various real-world benchmarks on existing harnesses without compromising accuracy
media — pass: The selected media is the verified direct link to the arXiv abstract page for this paper. This 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.
An AI system that chooses and carries out actions toward a goal, often using software tools.
Definition referenceA language model with many learned numerical settings, used to process and generate language.
Definition referenceA 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 arXivarXiv 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 referenceAn AI system that chooses and carries out actions toward a goal, often using software tools.
Definition referenceA language model with many learned numerical settings, used to process and generate language.
Definition referenceA 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 arXivarXiv 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