Salvato in:
| Autore principale: | |
|---|---|
| Natura: | Recurso digital |
| Lingua: | inglese |
| Pubblicazione: |
Zenodo
2025
|
| Soggetti: | |
| Accesso online: | https://doi.org/10.5281/zenodo.17677764 |
| Tags: |
Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
|
Sommario:
- <p>ABSTRACT<br>AI-assisted development promises substantial velocity gains, yet teams routinely fall short due to context loss and fragmented workflows. This paper introduces a two-layer context engineering architecture that treats context as a first-class system: (1) a declarative rule layer that encodes stable invariants, and (2) a programmatic Model Context Protocol (MCP) layer that exposes live project structure to AI agents. This architecture sits within a broader four-pillar synthesis consisting of an AI-optimized zero-friction stack, disciplined prompt workflows, the two-layer context system, and agentic orchestration patterns. We evaluate the approach through the development of a production-grade, multi-tenant SaaS platform (220k+ LoC) built by two part-time developers in fifteen weeks. Analysis of 3,676 AI-assisted sessions shows that Context Reuse Efficiency rose from 90.1% to 92.2% and Generative Amplification declined from 5.4% to 4.6%, consistent with a shift from generative substitution to retrieval-augmented reuse as context became systematized. Using a Design Science Research framework, we show that the two-layer pattern is feasible at production scale and produces measurable effects on development velocity. While based on a single case, the architecture and metrics establish a foundation for practitioners and researchers to validate and extend these practices across diverse contexts.</p>