Nicolas Gertler, Claudio Novelli and Luciano Floridi proposed a portable specification framework for agent harnesses in a preprint submitted to arXiv on 29 October 2026. The framework would record which model ran in which harness, treating the infrastructure surrounding a model as a versioned part of the system.[1]
The authors describe the harness as the infrastructure through which large language models act. Their abstract says they decompose it into eight functional components that can be specified and compared across systems. The saved abstract does not name those components or give the framework's detailed record format.[1]
The authors argue that different harness configurations can produce several-fold differences in reported scores for the same model, and that much of an agent's attack surface lies in its harness. These are claims in the abstract; our reading does not independently establish the score differences or assess the attack surface.[1]
Existing model and agent cards, they write, provide release and operational information without establishing the sequence of effective configurations during a run. Harnesses can change between releases, between runs and, when agents edit their own tools or configuration, within a run. The abstract also says enacted AI laws contain relevant documentation duties, subject to scope and application dates, without prescribing this common record.[1]
[1]: Towards Portable Specification of the Agent Harness. arXiv, 29 October 2026. https://arxiv.org/abs/2610.40917
Lead photo: Nicolas Gertler, Claudio Novelli, Luciano Floridi · CC BY 4.0.
Sources
Check or correct this page
Verification record
The digest below shows whether the text you are reading matches the published source. It does not show that the reporting is correct. The build computes it from this page's Markdown source and lists it in the site manifest.
7a7c3143296a7581ea257401ddcfb2e89ed00c75fa9a9b997fddf39b29703316
