단계별로 다른 모델을 사용하는 9개 에이전트 SDD 하네스 사용기: 월 10-15달러로 끝내는 전체 분석
I use a 9-agent SDD harness where each phase uses a different model. The total cost is $10-15/month. Here's the full breakdown.
핵심 요약
단계별 최적 모델을 배치한 9단계 SDD 워크플로우로 월 15달러 미만의 고효율 코딩 환경을 구축한 사례 분석.
- 단계별 모델 최적화 — 프로젝트 이해부터 아카이브까지 9단계마다 특화된 LLM을 배치해 비용과 성능을 모두 잡음.
- 저렴한 고정 비용 — OpenCode Go의 월 10달러 플랜을 활용해 DeepSeek, Kimi, GLM 등 주요 모델을 무제한급으로 사용함.
- 검증 및 오케스트레이션 — Qwen3-Coder로 구현물을 검증하고 Claude Sonnet 4.6으로 전체 워크플로우를 제어함.
- 실무 적용 사례 — .NET 8 마이크로서비스 환경에서 대규모 데이터를 다루는 아키텍트가 직접 구축한 실전형 워크플로우임.
Background: I'm a Principal Architect working on .NET 8 microservices at scale (~600 locations, 44k articles). I got tired of burning Claude/GPT tokens on tasks that don't need frontier reasoning, so I rebuilt my entire coding workflow around Spec-Driven Development with per-phase model selection.
The core insight is obvious once you see it: different phases need completely different capabilities . A phase that maps files has nothing in common with a phase that writes a formal spec. Running both on Claude Opus is like using a sledgehammer to hang a picture.
The 9-phase setup:
sdd-init → DeepSeek V4 Flash (OpenCode Go)
The goal of this phase is simply to build an initial understanding of the project. The agent maps the repository, detects conventions, identifies technologies, and gathers context that will be used throughout the workflow. There is very little reasoning involved at this stage. Speed and context size are far more important than deep analysis, which makes DeepSeek V4 Flash an excellent fit.
sdd-explore → Kimi K2.6 (OpenCode Go)
This is where the agent starts exploring the codebase in depth. It reads existing implementations, follows dependencies, analyzes test suites, and identifies patterns across the repository. Kimi performs particularly well here because it can process large amounts of information efficiently and leverage its agent capabilities to explore different parts of the codebase simultaneously.
sdd-propose → GLM-5.1 (OpenCode Go)
At this stage the objective is not to write code but to think through possible approaches. The agent evaluates alternatives, considers trade-offs, and proposes a direction before any implementation work begins. GLM-5.1 has proven especially strong at this kind of structured reasoning and technical decision-making.
sdd-spec → DeepSeek V4 Pro (High Reasoning)
The specification phase is one of the most important parts of the entire workflow. Every subsequent phase depends on the quality of the specification. If requirements are ambiguous or incomplete here, those problems will propagate into design, implementation, and verification. For that reason, this is one of the few stages where I always prioritize quality over cost.
sdd-design → DeepSeek V4 Pro (Medium Reasoning)
