Agentic Architecture for Software Development: A Framework for Context Engineering, Specs, and Agent Operations
The performance of AI coding agents depends not only on model capabilities but also on the architecture surrounding their operation, including context management, task specifications, execution policies, verification mechanisms, and observability. This whitepaper presents a conceptual and implementable framework for agentic software development, designed to improve decision quality, operational reliability, and governance while maintaining autonomy proportional to task risk. The proposed architecture introduces risk-based intent classification (I0–I3), structured context engineering, portable agent resources, evidence-based verification, independent review, and mechanisms for aligning implementation results with original requirements. The framework also distinguishes declared, observed, and independently verified evidence, addressing limitations in context injection visibility, policy enforcement, and agent autonomy across different coding environments. A reference implementation, Vostok AI Workflow, is presented through exploratory experiments and practical engineering observations. The results highlight both potential benefits and limitations, including governance overhead, context efficiency, verification reliability, and differences between agent execution environments. The reported measurements are preliminary, based on small samples, and have not been independently validated. The Vostok AI Workflow implementation and its underlying source code remain private. The paper contributes an architectural proposal for building more structured, auditable, portable, and risk-aware AI-assisted software development workflows. Keywords: Agentic AI, Software Engineering, Context Engineering, AI Coding Agents, Agent Governance, Spec-Driven Development, Software Architecture, Verification, Observability, Human-in-the-Loop, Vostok AI Workflow.
Authors
- Jesus Donaldo Sánchez Inzunza
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-19
- DOI
- https://doi.org/10.5281/zenodo.24361806
- Primary Topic
- Software Engineering Techniques and Practices
- Type
- preprint