adk optimize を試す - 2026/05/29 jagu'e'r LT

At Google Cloud Next '26, the new 'adk optimize' command was announced as part of the Agent Development Kit (ADK) to enable automatic prompt optimization. This presentation covers an overview of the command along with a real-world case study of how we used it to improve our prompts.

Google Cloud Next '26 にて、Agent Development Kit (ADK) の新機能として、プロンプトの自動改善を行う adk optimize コマンドが発表されました。
このプレゼンテーションではコマンドの解説と、それを利用して実際にプロンプトを改善した事例をご紹介します。

Transcript

1. adk optimize 2026/05/29 jagu'e'r LT
2. — — —
4. adk optimize — — —
5. 2023 — — 2025 . 07 — — — 2026 adk optimize — — — Lineage 03 / 15
6. 01 02 GEPA cores 04 / 15
7. CUSTOMER ↓ AGENT
8. instruction = ( " " " " )
9. 01 02 03 04 05 06 07 08
10. acknowledge_with_specifics action_with_timeline_and_owner hedge_uncertain_outcomes structured_three_part doesnt_take_bait urgency_match respects_choice — —
11. STEP 1 → STEP 2 → STEP 3 → STEP 4 → STEP 5 4 iterations
12. ~/adk-optimize-demo — zsh # 1. $ adk eval support_response support_response_eval \ --config_file_path=eval/test_config.json # 2. $ adk optimize support_response \ --sampler_config_file_path=eval/sampler_config.json \ --optimizer_config_file_path=eval/optimizer_config.json test_config.json rubrics: 4 acknowledge_with_specifics action_with_timeline_and_owner hedge_uncertain_outcomes structured_three_part ... eval sampler_config.json app_name: support_response train_eval_set: support_response_eval ... optimizer_config.json optimizer_model: gemini-2.5-flash reflection_minibatch_size: 3 max_metric_calls: 50 ...
13. BEFORE 4 / 8 → AFTER 7.67 / 8 Before 3 : 5 / 4 / 3 · After 3 : 8 / 8 / 7 LLM ( judge reflect) Gemini 2.5 Flash model — —
14. BEFORE 54 AFTER ( ) 1,483 1. 2. 3. ( ) 4. 1 ( ) 5. 6. SLA 7. ...
15. // REFLECTION-LM OUTPUT, ITER 0 → 1 doesnt_take_bait ( ) →