Long-Horizon Agents Need Experiments, Not Just Prompts — Erina Karati
記事はまだありませんこの動画の情報は登録済みですが、記事はまだ作成されていません。それまでは YouTube でご覧ください。YouTube で開く 同じチャンネルの記事
スライド 7 枚19:43EN
AI Engineer· 2026年10月26日Jack Wotherspoon on Building a Team of Managed AgentsJack Wotherspoon · Google DeepMindGoogle DeepMind’s Jack Wotherspoon introduces Gemini Managed Agents: cloud sandboxes that let teams run coding and research agents with a single API call. He explains how persistent environments, reusable skills, proxy-injected credentials, schedules, and specialized agents fit together.エージェント工学Agent orchestrationSandboxes & isolation+2
スライド 5 枚20:55ENAI Engineer· 2026年10月26日Jesse Hall：语音智能体要为对话而优化，而非只看排行榜Jesse Hall · LiveKitLiveKit 的 Jesse Hall 解释了为什么语音智能体必须像实时系统一样接受评测：听清用户、允许打断、在工具调用期间继续对话，并在用户真正听到声音之前控制延迟。音声 AIエージェント工学評価（Evals）+2 
スライド 12 枚17:56ENAI Engineer· 2026年10月25日How Bridgewater Keeps Its AI Analyst Inside Permission BoundariesNikhil Benesch, Aaron Linsky · turbopufferAaron Linsky and Nikhil Benesch explain why search over an enterprise vault must honor document permissions without sacrificing recall or speed. Their solution combines vector indexing with native filtering to search hundreds of millions of vectors in roughly 120 milliseconds.エンタープライズ AIPrivacy & data governanceSearch & retrieval+2 
スライド 7 枚35:19ENAI Engineer· 2026年10月25日SallyAnn DeLucia: How to Write Your First AI EvalSallyAnn DeLucia · Arize AITurn an agent’s real traces into concrete pass/fail criteria, then build an LLM judge you can test, align with people, and run in production.評価とオブザーバビリティ評価（Evals）エージェント工学+2 
スライド 9 枚20:58ENAI Engineer· 2026年10月25日Vinoo Ganesh on Making Financial AI Cite Every NumberVinoo Ganesh · KeplerKepler CEO Vinoo Ganesh explains why financial AI should retrieve and cite figures while deterministic software handles the math. His design makes traceability, retrieval quality, and tenant security part of the answer itself.金融 AIエンタープライズ AI検索拡張生成（RAG）+2 
スライド 10 枚18:31ENAI Engineer· 2026年10月25日Janie Gu and Suyeol Yun: Why Frontier Models Fail as Buy-Side AnalystsJanie Gu, Suyeol Yun · Google DeepMindJanie Gu explains why benchmark scores can mislead, while Suyeol Yun presents MandateBench, which tests whether a model can turn a fund mandate into a defensible investment memo. The two talks show why real workflows, reproducible grading, and evidence-backed reasoning matter.金融 AIBenchmarks評価（Evals）+2