Reproduce “A Contract-Centered Architecture for Scalable and Manageable Agentic Runtimes” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yaxiao Liu, Pengbo Liu, Yiwen Liu, Yihua Guan, Zhenghe Hou, Jiaxing Song
The authors propose contracts separating agent behavior, runtime governance, execution capacity, and data authority. The Skill defines versioned behavior, while the Harness governs activation and the Scaffold provides execution and isolation. Capability-capacity separability remains empirically untested. A reference model matched its oracle across 1,024 fault combinations for a revenue-record fixture. These checks establish conformance only, not retrieval quality, production safety, or scaling.
Key insights
- The paper defines cost-aware capability–capacity separability as a bounded interaction between business capability changes and infrastructure capacity changes, rather than incorrectly requiring capacity changes to have no performance effects.
- Complete mediation, source retention, atomic check-and-read operations, replayable contracts, and evidence envelopes are treated as executable controls rather than documentation conventions.
- The architecture distinguishes scientific evidence about task semantics from operational evidence about isolation, scheduling, resource behavior, and governance overhead.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.