Why Agentic Development Needs Policies: A Three-Level Governance Stack
Anthropic
OpenAI
Cursor
AI development tools are evolving from chat assistants to autonomous agents. This article describes a three-level governance stack built for a sports facility booking platform, arguing that policies become essential when agents make thousands of micro-decisions per hour. It details 12 policy documents, 10 skills, 6 specialized agents, and a 7-phase workflow.
The article argues that as AI coding evolves from chat assistants to autonomous agents, governance becomes a hard requirement. At level 1, a developer reviews every line; at level 3, agents make 10,000 micro-decisions per hour, making human review impossible. The author describes a governance stack for a platform using NestJS and Next.js, consisting of three layers: engineering policies (12 documents with numbered sections), skills and specialized agents (10 slash-commands, 6 agents, 6 review prompts), and a 7-phase workflow (EXPLORE, PROPOSE, DESIGN, SPEC, TASKS, APPLY, ARCHIVE) based on OpenSpec. Policies are contracts that agents reference mechanically, and stronger models review weaker ones. The workflow locks artifacts after review to avoid circular fixes. The article ends with an open question about the sustainability of governance in the context window as it scales.
- Сокращения
- API = Application Programming Interface — интерфейс программирования приложений
- CI = Continuous Integration — непрерывная интеграция
- TDD = Test-Driven Development — разработка через тестирование
- EARS = Easy Approach to Requirements Syntax — простой подход к синтаксису требований
- LCP = Largest Contentful Paint — время загрузки основного контента
- SQL = Structured Query Language — язык структурированных запросов
- SSH = Secure Shell — безопасный протокол удалённого доступа
- UX = User Experience — пользовательский опыт
Source: Habr — хаб ИИ —
original