Codex Workflow: Sol as Foreman, Luna Max as Boundary-Defined Worker
Category: tip · AI coding / Agent Harness Date: 2026-08-02 10:47 (Beijing time) Sources: OpenAI official docs and pricing notes; community workflow repository; AYi public hands-on testing
A Community Workflow Is Going Viral
On August 2, AYi shared a Codex usage pattern: the main thread, running GPT-5.6 Sol, handles task decomposition, architectural judgment, and final review; clearly scoped, verifiable implementation work, bug fixes, test runs, and refactoring go to GPT-5.6 Luna Max subagents. In short: Sol is the foreman, Luna is the worker.
This is not an official new product announcement, but a community-built workflow combining two existing capabilities. OpenAI's official documentation confirms that Codex supports parallel subagents and custom agent configurations. Agent files live in ~/.codex/agents/ or the project-level .codex/agents/, and must define name, description, and developer_instructions, with optional per-agent model and model_reasoning_effort settings.
Why This Layering Makes Economic Sense
OpenAI's July 30 pricing announcement provides firmer grounding: the GPT-5.6 Luna API price dropped to $0.20 per million input tokens and $1.20 per million output tokens; Sol pricing is unchanged. Luna can still call tools and complete multi-step workflows, making it well suited for handling large volumes of clearly bounded tasks.
OpenAI's official coding example follows nearly the same structure: Sol first handles uncertainty and defines the plan, then Luna implements well-defined changes, writes tests, runs them, and evaluates results. The community repository use-luna-subagents turns this idea into a reviewable configuration: showing a diff before installation, pinning the Luna model, preserving the parent session's sandbox and permissions, forbidding silent model substitution, and requiring the parent agent to verify results.
Don't Treat "Doubled Output" as a Benchmark Result
AYi's claim that "output volume on the same subscription directly doubled" is personal workflow experience, not a published controlled experiment. The official documentation instead highlights three easily overlooked points: subagents increase token consumption; permissions and sandbox settings are inherited from the parent agent by default; and parallel writes with overlapping scopes create conflicts.
A more reliable rollout approach: have Sol only delegate tasks that are rollback-safe, clearly bounded, and independently testable; do not let Luna directly modify shared configuration or critical interfaces; the main thread must review diffs, run tests, and then merge. Complex design, cross-module migrations, and security judgment remain Sol's job.
The significance of this development is not that "cheap models suddenly equal expensive models," but that the cost unit of AI coding is shifting from "one conversation" to "a set of layered agents plus a verification process." Model routing and harnessing are replacing simple model selection as the core of productivity.
Original sources and evidence:
- https://x.com/AYi_AInotes/status/2083867265179537565
- https://openai.com/index/advancing-the-price-performance-frontier-with-gpt-5-6
- https://developers.openai.com/codex/subagents
- https://github.com/aitransformationdirector/use-luna-subagents