7 Golden SubAgent Orchestration Rules to Let Expensive Models Only Talk and Cut Costs in Half

Problem: Expensive Model "Full‑Stack" Use Is Slow and Costly

Running a flagship large language model (LLM) to read thousands of lines of code, generate detailed solutions, and then modify code incurs high latency and token costs—often twice to twenty times the price of mid‑range models, while only about 10% of tokens cover valuable reasoning.

Core Idea: Let the Expensive Model Orchestrate, Cheap Models Execute

The solution splits a task into two roles: the main Agent (expensive model) handles requirement clarification, plan decomposition, task dispatch, and result verification with minimal token usage; SubAgents (cheaper models) perform all implementation work such as reading code, writing code, running tests, and batch modifications.

Seven General SubAgent Orchestration Rules (Applicable to All Model Combos)

Clarify before dispatching. SubAgents cannot ask follow‑up questions; ambiguous requests lead to costly rework.

Exempt tiny tasks. Simple edits or quick look‑ups should be done by the main Agent to avoid unnecessary orchestration overhead.

Dispatch independent tasks in parallel. Parallel SubAgents avoid the three‑fold time penalty of serial dispatch.

Prompts must be self‑contained. Each SubAgent prompt must include background, goal, file paths, constraints, and acceptance criteria.

Never trust the SubAgent’s report. The main Agent must read diffs, run tests, and verify key conclusions.

Specify model parameters explicitly. Use a directive like model: opus to ensure the intended cheap model is used.

Report after acceptance. Summarize what changed and how it was verified; users care only about the result, not the number of SubAgents.

Three Model Combinations and Complete Configurations

Combination A: Anthropic (Fable 5 + Opus)

Fable 5 costs about $10‑$50 per M tokens, roughly twice Opus’s $5‑$25 per M tokens, but excels at multi‑step reasoning. All implementation work is assigned to Opus, while Fable 5 is reserved for planning and verification.

## SubAgent 编排规则（Fable 5 + Opus）
- 先澄清后派发：subagent 无法向用户提问，需求不明确时先问用户，确认后再派发。
- 实现类工作派给 Opus：读大量代码、写代码、跑测试、批量修改等，用 Agent 工具派发并指定 model: opus。
- 豁免：单文件小改动、快速查询等低成本操作，自己直接做。
- 并行：相互独立的任务并行派发多个 Opus subagent。
- 派发要求：每个 subagent prompt 必须自包含——背景与目标、相关文件路径与关键位置、约束条件、明确的完成标准。
- 验收：不轻信 subagent 汇报。读 diff、跑测试/构建验证，关键结论自己核查；不合格则带具体反馈重新派发。
- 汇报：验收通过后向用户简要汇总改了什么、如何验证的。

Combination B: GPT‑5.6 Three‑Tier (Sol / Terra / Luna)

Sol ($30 /M) is the most expensive but strongest planner; Terra ($2.5‑$15 /M) handles coding; Luna ($1‑$6 /M) handles batch‑type mechanical work.

## SubAgent 编排规则（GPT-5.6：Sol / Terra / Luna）
- 先澄清后派发：subagent 无法向用户提问，需求不明确时先问用户，确认后再派发。
- 分级派发：编码实现类 → Terra (model: terra)；机械批量类 → Luna (model: luna)。
- 豁免：单文件小改动、快速查询等低成本操作，自己直接做。
- 并行：相互独立的任务并行派发多个 subagent。
- 派发要求：每个 subagent prompt 必须自包含——背景与目标、相关文件路径与关键位置、约束条件、明确的完成标准。
- 验收：不轻信汇报。读 diff、跑测试/构建验证，关键结论自己核查；不合格则带具体反馈重新派发。
- 汇报：验收通过后向用户简要汇总改了什么、如何验证的。

Combination C: Domestic (Qwen3.8‑Max + DeepSeek‑V4‑Flash)

Qwen 3.8‑Max’s output price is >20× DeepSeek‑V4‑Flash’s. Qwen acts as the planner; all implementation work is delegated to Flash.

## SubAgent 编排规则（Qwen3.8-Max + DeepSeek-V4-Flash）
- 先澄清后派发：subagent 无法向用户提问，需求不明确时先问用户，确认后再派发。
- 实现类工作派给 DeepSeek‑V4‑Flash：读大量代码、写代码、跑测试、批量修改等，用 Agent 工具派发并指定模型。
- 豁免：单文件小改动、快速查询等低成本操作，自己直接做。
- 并行：相互独立的任务并行派发多个 subagent（DeepSeek 支持高并发）。
- 派发要求：每个 subagent prompt 必须自包含——背景与目标、相关文件路径与关键位置、约束条件、明确的完成标准；项目背景前缀保持稳定，最大化缓存命中（缓存价仅为普通输入价的 1/50）。
- 格式要求写明确：Flash 在严格格式约束下偶有越界，派发时格式说明要具体，验收时逐项核对结构化输出。
- 验收：不轻信汇报。读 diff、跑测试/构建验证，关键结论自己核查；不合格直接带反馈重派（Flash 试错成本极低）。多模态验收（截图、图表、界面审查）由 Qwen 主 Agent 自己完成。
- 峰谷调度：DeepSeek 自 2026‑08‑17 起分峰谷计价（高峰约为闲时两倍），非紧急的大批量任务尽量排在闲时执行。
- 汇报：验收通过后向用户简要汇总改了什么、如何验证的。

Practical Test and Acceptance Workflow

Open a new session. Ensure the global instruction file is not re‑loaded automatically.

Submit a medium‑scale task. Example: 新增格式化字节函数 format_byte_size 并补测试.

Verify execution according to the rules. Two methods:

Ask the main model for a summary of dispatches (quick sanity check).

Inspect backend logs to see which model handled each request and the associated cost. Example commands:

# Find the latest rollout file
f=$(find ~/.codex/sessions -name "rollout-*.jsonl" | tail -1)
# Count messages per model
grep -o '"model":"[^"]*"' "$f" | sort | uniq -c
# Show model order
grep -o '"model":"[^"]*"' "$f" | uniq

Acceptance checklist. Confirm: (1) clarification before dispatch, (2) cheap model specified for implementation, (3) tiny tasks handled by main Agent, (4) independent tasks dispatched in parallel, (5) manual diff/read and test verification, (6) concise report to the user.

Six Common Pitfalls

Writing conditional branches based on the model name—models cannot recognize their own identifiers.

Dispatching every task without a lower‑bound—overhead can exceed the task cost.

Omitting context in SubAgent prompts—leads to random or incorrect outputs.

Serial dispatch of independent tasks—wastes the parallelism advantage.

Accepting SubAgent reports without diff/read and test verification.

Dispatching before clarification—ambiguous requests cause full‑cycle rework.

Final Takeaway

The essence of orchestration mirrors human team management: the most expensive resource should make decisions, not perform repetitive work. The larger the price gap between models, the greater the savings—e.g., a 20× price difference in the domestic combo can deliver ten times more work for the same budget. Consistently applying the three core habits—clarify first, use self‑contained prompts, and verify personally—will thin the bill and speed up delivery.