VS CodeのCodexでLuna / Terra / Solを自動で使い分ける ― AGENTS.mdとSubagentでモデルルーティングを作るThis article explains how to build a model-routing setup in VS Code Codex that…

Developers working with AI coding assistants increasingly face a trade-off between capability and cost: the most powerful models tend to produce better results but charge more, while lighter models are cheaper yet can stumble on complex work. A recent walkthrough published on Qiita describes a practical way to manage this inside VS Code Codex, using configuration files and delegation features to automatically route tasks to one of three models — Luna, Terra, and Sol — depending on how demanding each request is.

The central concept is model routing. Rather than manually switching models for every request, the setup lets Codex select which model to invoke based on task complexity. Simple edits, formatting, or boilerplate generation can be handled by a lighter model, while architectural reasoning, multi-file refactoring, or debugging that spans a codebase can be escalated to a heavier one. The article frames this as a way to balance cost management with processing efficiency, since the amount billed under pay-as-you-go pricing changes depending on which model ends up doing the work.

Two mechanisms make the routing possible. The first is AGENTS.md, a markdown file placed in a project that gives an AI agent standing instructions about how to behave. In this configuration, AGENTS.md carries the routing rules — the guidance that tells the agent when to reach for Luna, Terra, or Sol. The second is the Subagent feature, which lets a primary agent spin up specialized sub-agents to handle delegated portions of a task. By pairing routing rules with subagents, the main agent can hand off subtasks to whichever model suits them, keeping the more expensive calls reserved for the parts that genuinely need them.

The article is explicit that it reflects the state of Codex and the GPT-5.6 series as of August 25, 2026, and it cautions that model names, configuration options, and the Subagent specification may change. That disclaimer matters because tooling in this space has moved quickly, and a routing setup that works today could require adjustment as models are renamed, retired, or repriced. Readers are advised to treat the configuration as a starting point and to verify current behavior against their own environment, and the author notes that use of the described approach is at the reader's own risk.

Some background helps explain why this pattern is gaining traction. AGENTS.md has emerged as a loose convention across several coding-agent tools for supplying repository-specific instructions, similar in spirit to how a README orients human contributors or how linter configuration standardizes code style. The Subagent, or sub-agent, pattern reflects a broader industry shift toward orchestration — decomposing a large request into smaller, specialized units of work rather than relying on a single monolithic prompt. Comparable ideas appear in other agent frameworks that route between models or split reasoning and execution across components, so the technique described here is best understood as one implementation of a wider trend rather than a wholly novel invention.

For teams, the appeal is straightforward. If most routine coding tasks can be served by a cheaper model, the aggregate bill under usage-based pricing can fall while a stronger model remains available for the minority of hard problems. The naming of Luna, Terra, and Sol appears to correspond to distinct tiers within the GPT-5.6 lineup, each with its own capability and price profile, though the exact characteristics are subject to the same caveat about changing specifications.

The trade-off is added configuration complexity and the need to tune routing rules carefully. If the rules are poorly calibrated, tasks may be under-served by a model that is too weak, forcing retries, or over-charged by one that is stronger than necessary. In principle, aggressive escalation could even raise costs rather than lower them, so the value of the approach depends heavily on how well the thresholds match a team's actual workload.

For those already using Codex in VS Code, the walkthrough offers a concrete template for combining declarative instructions in AGENTS.md with delegated execution through subagents. As with any usage-based AI tooling, the practical advice is to start small, monitor billing closely, and revisit the routing logic whenever the underlying models or their pricing change. The concept is likely to remain relevant even if the specific model names do not, given the persistent tension between quality and cost that motivates it.