はじめに
㊗️アイうた5周年!nikkie (UUID 28fb3f96-a221-462c-93bd-567b431715b9) です。
10/29(木)AM1:30〜2:30(JST)に行われたADK Community Callで知った事項を残します。
睡魔と戦いながら参加しました(翌日あるので途中離脱)。
9月の参加はこちら
目次
アジェンダ
予告のメールより
https://groups.google.com/g/adk-community/c/jSpzT9NdW60
- ModelConsultTool (Mid-Task Escalation): Keep bulk execution fast and cheap on Gemini Flash while consulting Gemini Pro as an advisor only on tricky edge cases
- Skill Lifecycle Management: Prevent context bloat and leftover tools in long sessions with new BOUNDED (LRU), EPHEMERAL, and on-demand unload_skill modes.
- Classifier API & Jev (~170ms Agent Decisions): Route workflows and gate high-risk tool calls 4–12x faster than standard LLM prompts, featuring a live side-by-side browser agent demo.
- How Gemini 4 Argon Sped Up ADK by 40%: What happened when we gave Argon one /goal prompt and our .agents/skills review rules
ADK Community Callはアーカイブのplaylistがあるので、少し待てば公開されると思います(でも9月がまだ入ってないような...)
v2.11.0で追加
ModelConsultTool: Let an agent consult another model mid-task, capped by per-turn and per-session budgets. See the unit guide for more details. (84cc99a)
adk-python内のunit guide
https://github.com/google/adk-python/blob/v2.13.0/docs/guides/tools/model_consult/model_consult_tool/index.md
サンプルより1
root_agent = Agent(
name='order_support_agent',
description='Handles customer order returns, warranty defect bulletins, and loyalty refund policies.',
instruction=_TASK_INSTRUCTION,
tools=[
get_order,
get_customer_profile,
issue_refund,
ModelConsultTool(
max_uses=2,
session_max_uses=5,
thinking_level='high',
),
],
)
Skill Lifecycle Management
ADKはv1.25.0でAgent Skillsをサポートしていますが、v2.10.0でそのコンテキスト管理(Lifecycle Management)を追加!
Skills: Control resource usage and tool persistence dynamically using new ephemeral lifecycles and active skill limits (experimental; enable with ADK_ENABLE_SKILL_LIFECYCLE=1). (43b73e4)
ガイドからControl how long skills stay active
The PERSISTENT mode is the default. The skill stays active for the rest of the session.
The BOUNDED mode works like PERSISTENT, except that at most max_active_skills bounded skills can be active for an agent at once.
The EPHEMERAL mode keeps the skill active for the rest of the turn that loaded it and releases it when the next turn begins.
toolset = SkillToolset(
skills=skills,
lifecycle_config=SkillLifecycleConfig(
default_mode=SkillLifecycleMode.BOUNDED,
max_active_skills=3,
skill_overrides={"incident-report": SkillLifecycleMode.EPHEMERAL},
),
)
skill(特定の専門知識)をコンテキストに入れたまま何十ターンも連れ回すのか? その必要がない場合の選択肢が入ったのは個人的に好感触です。
Classifier API & Jev
大バズりのSystem One ModelにもADKは飛びついていきます!(v2.13.0)
Classifiers: Route workflows with TypeSafe's Jev models via the new TypesafeClassifier and ClassifierNode. Jev returns calibrated probabilities in a single forward pass, much faster than an LLM. See the unit guide for more details. (ddf8c55, e8c4bac)
unit guideは2つ
サンプルも2つ
(1) ADK Workflow Classifier Routing Sample
Jevでない例です
classifier = LlmClassifier(model="gemini-3.5-flash")
triage = ClassifierNode(
name="triage",
classifier=classifier,
question=Choice(
instructions="Which team should handle this support ticket?",
criteria={
"billing": "Invoices, refunds, or subscription charges",
"technical": "Bugs, errors, outages, or API problems",
"general": "Product questions and anything else",
},
),
answer_key="triage_answer",
)
(2) Classifier Tool Gate Sample
before_tool_callbackに渡す関数でLlmClassifierを使います。
https://github.com/google/adk-python/blob/v2.13.0/contributing/samples/hitl/classifier_tool_gate/tool_gate.py#L50
Before each tool call, the callback asks classifier a Binary question about the call. If the probability of true is at or above threshold, the call is blocked, or held until the user confirms it.
コーディングエージェントのauto mode(代理で承認モード)がADKのエージェントで実現できていそうです。
How Gemini 4 Argon Sped Up ADK by 40%
v2.12.0でArgonによる改善が入っています
Performance: Cut time to first model request by 40% in long sessions. (1754b3e)
コミットメッセージ抜粋
before after
first model request after 50 turns * 64 ms 36 ms -43%
first model request after 150 turns * 200 ms 122 ms -39%
adk-pythonにはAgent Skillsが整備されており、
https://github.com/google/adk-python/tree/main/.agents/skills
Argonにそれらを装備してうまくいったという報告です。
With .agents/skills in the repo and /goal mode active, Argon ran unattended for 11 hours overnight (Figure 3).(Googleブログより)
終わりに
10月のADK Community Callで知った情報を書き出しました。
ModelConsultTool、SkillLifecycleConfig、LlmClassifier、これらを使うことでこれまでできなかったエージェントの開発ができそうと可能性を感じています。
そしてこの記事では取り上げきれませんでしたが、実は他にもリリースがあります(YouTubeアーカイブ出てくれ!)
