Not yet human-reviewed
This source record was written automatically by the research pipeline and has not yet been reviewed by the maintainer or integrated into a topic page. Treat it as a provisional summary and verify claims against the original source.
A Contract-Centered Architecture for Scalable and Manageable Agentic Runtimes
Citation. Liu, Y., Liu, P., Liu, Y., Guan, Y., Hou, Z., & Song, J. (2026). A Contract-Centered Architecture for Scalable and Manageable Agentic Runtimes. arXiv:2608.27086.
Summary. This conceptual preprint proposes a contract-centered architecture for enterprise agentic runtimes that separates versioned business capabilities, runtime governance, execution capacity, and a stack-external enterprise data substrate. It frames those separations as shared organizational contracts over ownership, change cadence, and evidence obligations. The paper specifies a falsifiable capability-capacity separability hypothesis and a proposed cluster-period randomized crossover evaluation, but reports no completed implementation, experiment, dataset, or measured result.
Key claims
- The proposed architecture distinguishes a Skill that owns bounded business semantics, a Harness that governs admission and composition, a Scaffold that owns execution and non-functional requirements, and an independently governed data substrate.
- The paper argues that enterprise agentic systems should test whether capability changes preserve capacity-response behavior and whether compatible capacity changes preserve capability semantics, rather than inferring scalability from task success, throughput, or latency alone.
- The proposed evaluation treats policy enforcement, effect authorization, evidence capture, and versioned contracts as measurable obligations whose coverage, violations, uncertainty, and cost determine whether the separability claim is decidable.
Evidence & limitations
- This is an arXiv preprint and a conceptual architecture plus falsification methodology; it reports no completed runtime implementation, experimental dataset, or measured results.
- The public arXiv v1 PDF exactly matches the ingested Drive artifact by SHA-256; arXiv currently lists a later v2 record, so this record describes the ingested v1 text and uses the stable unversioned arXiv landing page.