Gertler, Nicolas, Novelli, Claudio, Floridi, Luciano · Oct 29, 2026
Three authors propose a portable specification for the agent harness
Markdown · JSON · Original source
Nicolas Gertler, Claudio Novelli and Luciano Floridi argue that the infrastructure surrounding a —the harness—shapes what the model can do and how it fails, yet lacks a settled definition. They propose breaking the harness into eight functional components and recording, in a portable specification, which model ran in which harness. Their concern is that different harness configurations can produce several-fold differences in reported scores for the same model, while much of an agent’s attack surface lies in its harness. Model cards and agent cards provide release and operational information, the authors say, but do not record of effective configurations within a run. That sequence matters to their proposal because harnesses may change between releases, between runs, or within a run when agents edit their own tools and configuration. The authors distinguish their proposal from enacted AI laws: those laws include documentation duties for covered systems, subject to their respective scopes and application dates, but do not prescribe the common harness record they propose. They conclude that the harness should be explicitly described and specified as a versioned part of the system.
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: Oct 29, 2026 (2026-10-29).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-10T10:54:58.370Z. First observed: 2026-11-02T05:05:03.049Z. Retrieved: 2026-11-02T07:49:55.533Z.
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, not its full text; it does not describe the eight functional components or the details of the portable specification framework. The abstract does not explain how the proposed record would capture configuration changes made within a run.
Open question
How would the proposed framework record of effective configurations when an agent changes its tools or configuration within a run?
Source-stated
The authors say the infrastructure surrounding a large language model, called the harness, shapes what it can accomplish and how it fails, but lacks a settled definition.
That infrastructure, the harness, determines what a model can accomplish and how it fails, yet it has no settled definition.
Source-stated
The authors propose decomposing the harness into eight functional components and a portable specification framework recording which model ran in which harness.
we decompose the harness into eight functional components that can be specified and compared across systems, and we propose a portable specification framework that records which model ran in which harness.
Source-stated
The authors say different harness configurations can produce several-fold differences in reported scores for the same model, and much of an agent's attack surface lies in its harness.
different harness configurations can yield several-fold differences in reported scores for the same model, and because much of an agent's attack surface lies in its harness.
Source-stated
The authors say model cards and agent cards do not establish the sequence of effective configurations within a run; harnesses may change between releases, runs, or within a run.
do not establish the sequence of effective configurations within a run. Harnesses can change between releases and between runs, and, where agents edit their own tools and configuration, within a run.
Source-stated
The authors say enacted AI laws contain relevant documentation duties for covered systems but do not prescribe their proposed common harness record; they conclude the harness should be specified as a versioned part of the system.
Enacted AI laws contain documentation duties relevant to covered systems, subject to their respective scopes and application dates, but do not prescribe the common harness record proposed here. Therefore, the harness has to be explicitly described and specified as a versioned part of the system.
Source and contributor context
Gertler, Nicolas, Novelli, Claudio, Floridi, Luciano · 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.
Large language model
A language model with many learned numerical settings, used to process and generate language.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
The Sequence
The Sequence publishes machine-learning analysis and interviews whose guest answers retain their own attribution.
Editorial provenance
Version 1. Publication ID: c331c5c4-0cac-4f19-895b-bd01387a6f53.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: passed by anthropic/claude-opus-5.5 on 2026-11-10T10:54:58.370Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: useful_perspective.
Observation: e4fed5d03bead5a4a54bac2ca96ccbef410cde437baaab944bd12352ad19490c. Source text hash: 0bcdb9c69a34a060472151b81e88dfa922eb875725b377cbad1a26c845d224cf.
Revision and source-check history
Version 1 · 2026-11-10T10:54:58.370Z
Three authors propose a portable specification for the agent harness
Nicolas Gertler, Claudio Novelli and Luciano Floridi argue that the infrastructure surrounding a —the harness—shapes what the model can do and how it fails, yet lacks a settled definition. They propose breaking the harness into eight functional components and recording, in a portable specification, which model ran in which harness. Their concern is that different harness configurations can produce several-fold differences in reported scores for the same model, while much of an agent’s attack surface lies in its harness. Model cards and agent cards provide release and operational information, the authors say, but do not record of effective configurations within a run. That sequence matters to their proposal because harnesses may change between releases, between runs, or within a run when agents edit their own tools and configuration. The authors distinguish their proposal from enacted AI laws: those laws include documentation duties for covered systems, subject to their respective scopes and application dates, but do not prescribe the common harness record they propose. They conclude that the harness should be explicitly described and specified as a versioned part of the system.
Original source
cs.AI · Article
Source details
Original source date: Oct 29, 2026 (2026-10-29).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-10T10:54:58.370Z. First observed: 2026-11-02T05:05:03.049Z. Retrieved: 2026-11-02T07:49:55.533Z.
Limits
The supplied evidence is the paper's abstract, not its full text; it does not describe the eight functional components or the details of the portable specification framework. The abstract does not explain how the proposed record would capture configuration changes made within a run.
Open question
How would the proposed framework record of effective configurations when an agent changes its tools or configuration within a run?
Source-stated
The authors say the infrastructure surrounding a large language model, called the harness, shapes what it can accomplish and how it fails, but lacks a settled definition.
That infrastructure, the harness, determines what a model can accomplish and how it fails, yet it has no settled definition.
Source-stated
The authors propose decomposing the harness into eight functional components and a portable specification framework recording which model ran in which harness.
we decompose the harness into eight functional components that can be specified and compared across systems, and we propose a portable specification framework that records which model ran in which harness.
Source-stated
The authors say different harness configurations can produce several-fold differences in reported scores for the same model, and much of an agent's attack surface lies in its harness.
different harness configurations can yield several-fold differences in reported scores for the same model, and because much of an agent's attack surface lies in its harness.
Source-stated
The authors say model cards and agent cards do not establish the sequence of effective configurations within a run; harnesses may change between releases, runs, or within a run.
do not establish the sequence of effective configurations within a run. Harnesses can change between releases and between runs, and, where agents edit their own tools and configuration, within a run.
Source-stated
The authors say enacted AI laws contain relevant documentation duties for covered systems but do not prescribe their proposed common harness record; they conclude the harness should be specified as a versioned part of the system.
Enacted AI laws contain documentation duties relevant to covered systems, subject to their respective scopes and application dates, but do not prescribe the common harness record proposed here. Therefore, the harness has to be explicitly described and specified as a versioned part of the system.
Source and contributor context
Gertler, Nicolas, Novelli, Claudio, Floridi, Luciano · 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.
Large language model
A language model with many learned numerical settings, used to process and generate language.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
The Sequence
The Sequence publishes machine-learning analysis and interviews whose guest answers retain their own attribution.
Editorial provenance
Version 1. Publication ID: c331c5c4-0cac-4f19-895b-bd01387a6f53.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: passed by anthropic/claude-opus-5.5 on 2026-11-10T10:54:58.370Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: useful_perspective.
Observation: e4fed5d03bead5a4a54bac2ca96ccbef410cde437baaab944bd12352ad19490c. Source text hash: 0bcdb9c69a34a060472151b81e88dfa922eb875725b377cbad1a26c845d224cf.
passed · 2026-11-10T10:54:58.370Z · anthropic/claude-opus-5.5
Publication: c331c5c4-0cac-4f19-895b-bd01387a6f53
attribution — pass: The reading correctly credits Nicolas Gertler, Claudio Novelli and Luciano Floridi, as listed on the arXiv page. It keeps the abstract's assertions, including the score differences, attack surface and legal points, attributed to the authors rather than presenting them as established fact.
Authors:Nicolas Gertler, Claudio Novelli, Luciano Floridi
summary — pass: The title, summary and question faithfully follow the abstract's argument and conclusion. The harness is defined in plain language, and the reasons the authors give are clearly stated. Terms such as attack surface and model cards could use term-library tooltips, but they do not block understanding. The question addresses a real gap that the abstract leaves open.
Therefore, the harness has to be explicitly described and specified as a versioned part of the system.
date — pass: The originalPublished date of 2026-10-29 matches the v1 submission date shown on arXiv. The later feed entry date of 2026-11-02 is not used.
[Submitted on 29 Oct 2026]
support — pass: Every claim is source-stated and carries an exact supporting excerpt. Using 'shapes' for the abstract's 'determines' is a slight softening but not a distortion. The reading adds no invented benefits, costs or forecasts. The limitations correctly note that only the abstract is available and that the eight components are not described.
different harness configurations can yield several-fold differences in reported scores for the same model
media — pass: The selected media is the verified direct link to the original arXiv abstract page. It is the most direct way for readers to access 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.
A language model with many learned numerical settings, used to process and generate language.
Definition referenceThe Sequence publishes machine-learning analysis and interviews whose guest answers retain their own attribution.
Explore The SequencearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivThe Sequence publishes machine-learning analysis and interviews whose guest answers retain their own attribution.
Explore The SequenceA language model with many learned numerical settings, used to process and generate language.
Definition referenceThe Sequence publishes machine-learning analysis and interviews whose guest answers retain their own attribution.
Explore The SequencearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivThe Sequence publishes machine-learning analysis and interviews whose guest answers retain their own attribution.
Explore The Sequence