Stop Using One Frontier Model for Everything: My Role-Based AI Coding Stack
The current AI coding discussion spends too much time asking which model is “best.” I think that framing is already becoming obsolete. There may be a best model for a particular benchmark, but software engineering is not one task. It is a pipeline containing planning, repository exploration, implementation, debugging, documentation, testing, validation and adversarial review. Why should every…
The current AI coding discussion often asks which model is "best." However, this framing is becoming obsolete because software engineering is a pipeline comprising planning, repository exploration, implementation, debugging, documentation, testing, validation, and adversarial review. It is clear that not every stage should use the same model.
The author proposes treating models like members of an engineering organization, with specific roles such as ORCHESTRATOR Sol, PLANNER Opus, WORKERS (Gemini Flash, DeepSeek, GLM, Luna, Muse, MiniMax), ESCALATION Sonnet, Qwen 3.8 Max, and Terra, and TRUTH SEEKERS Grok and Kimi.
Workers are optimized for throughput and receive bounded tasks like implementing a validated plan, adding tests, refactoring code, tracing exceptions, generating fixtures, and inspecting modules. Escalation involves using stronger models like Sonnet, Qwen 3.8 Max, and Terra when workers encounter issues that require greater reasoning.
Planning should be a separate capability because a bad plan can have significant consequences. The author prefers using Opus for planning, as it establishes objective constraints, affected components, architecture dependencies, implementation sequence, acceptance criteria, rollback considerations, and verification strategy. Orchestration, on the other hand, determines what happens next based on system state, task dependencies, and execution outcomes.
The author also emphasizes the need for a truth-seeking lane for adversarial review, with models like Grok and Kimi serving as independent reviewers. They attack the results and look for incorrect assumptions, architectural violations, security vulnerabilities, concurrency problems, and missing tests. Using different model families for different tasks introduces useful disagreement and prevents reliance on blind spots in a single model.
Finally, the author stresses that specialists still matter for tasks like documentation, generating synthetic datasets, and producing evaluation cases. These workloads can often be delegated without needing the "best model" approach, saving resources and improving efficiency.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
