Agent SDLC on an Internal LLM: Engineering Instead of Prompts
OpenCode
In a corporate environment with strict security policies, external LLMs are prohibited, and internal models are much weaker. The article describes an approach using structural engineering—deterministic code blocks, lazy-loaded rules, and subagents—to reliably complete real tasks.
The article presents an agentic SDLC approach developed for a large Android project using an internal LLM gateway (no frontier models allowed). The method relies on structural engineering rather than prompts: rules are stored in tables and read fresh when needed, agents are isolated via files, and dangerous actions are blocked by deterministic code. The harness includes lazy-loaded rules, subagents for context isolation and role specialization, vision-to-text routing via a vision model, and a three-level protection system (prompts, permissions, plugins). The orchestration command /ticket follows a spec-driven development flow: triage, workspace setup, design extraction, analysis, round-trip questions, test plan, implementation, and spot-check. Each agent has a narrow role, and context is managed carefully to avoid wasting tokens. Several incidents from the work log illustrate how structural fixes prevented recurrences of errors.
- Сокращения
- SDLC = Software Development Life Cycle — жизненный цикл разработки ПО
- LLM = Large Language Model — большая языковая модель
- CLI = Command Line Interface — интерфейс командной строки
- MCP = Model Context Protocol — протокол контекста модели
Source: Habr — хаб ИИ —
original