OpenAI Cookbook：客服 agent 成本與品質最佳化，四輪把每張工單成本砍 95%
Optimizing Customer Support Agents for Cost and Quality
摘要分析
OpenAI Cookbook（repo ⭐7.7 萬）2026-09-14 驗證的這篇教一套可重複的最佳化迴圈：量基準、每次只改一件事、品質過關才接受省下的成本。範例是電商客服 agent，四輪結果：壞基準每張工單 11,935 token、$0.038；收緊提示與工具 → $0.005 且品質 0.51→0.98；模型路由 → $0.003；prompt caching → $0.0024；QA 搬到背景 Batch → $0.002。10 萬張工單從 $3,813 降到 $204，約省 95%。
- 先量基準（品質、token/工單、延遲、成本），每輪只改一件事
- 第 1 招：精確任務規則、低冗長度、輸出上限、每次只給必要工具
- 第 2 招：模型路由，簡單步驟用 mini/nano，高風險才用大模型
- 第 3 招：prompt caching，穩定的指令與政策放最前、工單資料放最後
- 第 4 招：QA、標記、稽核搬到背景或 Batch API；程式碼可離線跑，設 RUN_LIVE_API_CALLS=true 才打 API
Verified on 2026-09-14, this OpenAI Cookbook recipe (repo about 76.6k stars) teaches a repeatable optimization loop: measure a baseline, change one thing at a time, and accept savings only after quality gates pass. The example is a fictional e-commerce support agent with five mock tools and ten synthetic tickets. Round results: the bad baseline used 11,935 tokens and $0.038 per ticket at 0.51 quality; round 1 (specific rules, low verbosity, output caps, limited tool set) dropped to 1,379 tokens and $0.005 at 0.98 quality; round 2 model routing (small models for triage, full model for risky cases) reached $0.003; round 3 prompt caching (stable instructions first, ticket data last) $0.0024; round 4 moving QA and analytics to background Batch processing $0.002. At 100k tickets that is $3,813 versus $204, roughly a 95% cut, though the authors stress it is a modeled simulation.
- Measure a baseline, change one thing per round
- Round 1: tight rules, low verbosity, output caps, limited tools
- Round 2: route simple steps to mini/nano models
- Round 3: prompt caching with stable content first
- Round 4: move QA and analytics to Batch
關鍵字：OpenAI Cookbook成本最佳化prompt caching模型路由Batch API
來源原文：https://developers.openai.com/cookbook/examples/agent_optimization/optimizing_agents_for_cost_and_quality.md
本頁為報導內容的摘要與分析，版權屬原媒體所有；請以原文為準。