Where Builders
Ship AI Ideas
That Scale

過去幾年，AI 輔助開發多半以短時間的互動為主：工程師提出需求，模型給出回應，人再接手檢查與整合。然而，隨著 Agent 能夠自主執行數小時的任務，這種分工模式正在改變。本場演講將從軟體開發生命週期（SDLC）的角度出發，探討長時間運作的 Agent 會如何影響需求分析、系統設計、開發、測試、維護等各個階段。

This keynote, titled "The Dawn of the Agentic Era," sets global technology trends as its macro backdrop—spanning key perspectives from the Sequoia Capital AI Summit, definitions of AGI and the divergence across industries, the timeline of the cognitive revolution, and the stratification of moats and opportunities. It then dives into seven practical case studies across three themes:

• Industry practice in science and finance: autonomous research replication in the scientific domain, and a multi-agent system for quantitative trading backtesting in finance;
• Agentic applications for creativity and productivity: turning a blog post into a slide deck in three minutes (Kiro), generating 3D designs from a product image (GPT-6 Astra on Bedrock), end-to-end indie game development driven entirely by AI (GPT-6 Astra on Bedrock), and using the Agent Toolkit for AWS (with Codex and Claude Code) to help developers build and ship to production faster on AWS;
• Deployment and cost optimization in practice: deploying an open-source model on AWS (Flux text-to-image on EC2) to meaningfully reduce high token costs.

Finally, the speaker shares the methods and results of the past three months—using a multi-agent workflow (Obsidian + multi-model agents) to achieve a 10× leap in personal productivity—and calls on developers to embrace this new way of working as soon as possible.

從用生成式 AI 從零做出「AI 名片王」，到一個人每天產出多支影片的內容工作流，分享 AI agent 實際落地時的問題、雲端與 API 成本怎麼算，以及 AI 降低開發門檻之後，工程師與創作者的價值會移到哪裡。本場採一對一對談，也會回應學生與跨領域聽眾怎麼開始動手。

1. 用OSI七層模型理解DDoS攻擊型態
2. AWS如何對應不同層級的防護需求
3. 從架構角度建立分層防禦思維

一個簡單的網站，真的需要複雜的雲端架構嗎？這場分享將以實際網站為例，透過架構圖快速看懂 Serverless 架構，帶大家看看如何運用 Kiro 協助開發，搭配 AWS Serverless 服務，打造一個簡單、輕量又容易維護的網站

當多人共用同一個 AI Agent，例如：不同使用者透過同一 Slack Bot 存取 GitHub 時，常見做法是讓 AI Agent 使用一組共用帳號或 credential，但這會造成不同使用者的操作都顯示為同一個帳號，帶來權限過大、難以追溯原始使用者、權限控管困難，以及難以落實 Least Privilege 等問題。

這場分享將從一個「多人共用同一 AI Agent 存取第三方服務」 的情境出發，了解共用 AI Agent credential 帶來的 Security 與 Audit 問題，接著介紹如何使用 AgentCore Gateway 與 AgentCore Identity，讓 AI Agent 代表原始使用者，使用該使用者授權的身份存取第三方服務。

透過這種方式，可以沿用第三方軟體等既有系統的權限控管、維持原本的 Authorization Boundary，並讓每次 AI Agent 操作都能追溯到真正的使用者。

An SRE's day often starts with an alert and a wall of error logs: read the logs, guess the cause, check whether we've seen it before. In this session I show how I turned that routine into a loop that improves over time with Victoria Gateway, an open-source project I built: AI reads the logs first and suggests a fix, engineers confirm it, and the case goes into a RAG knowledge base shared across on-prem and AWS.

LLM 接上護欄（Guardrails）後怎麼知道它有沒有生效？本場用 Amazon Bedrock 合成出測試資料、用於驗證 Amazon Bedrock Guardrails；並以金融資訊助理為例驗證一般金融問題能不能正常回答、而超過任務範圍的請求會不會被攔下。

也會分享如何檢查測資、找出誤擋與漏擋，再調整及複測護欄是否真的有效。

一、十四招最佳實務：包含Optimized Reads、Aurora I/O 優化、參數優化等機制。在處理單表超過 1300G、近 28 億筆資料與每日大量高壓讀下，展現出優異的系統穩定度。

二、成本優化：Aurora Serverless 在一天中有 22 小時維持低規格運作、將磁碟空間大幅精簡，以多項最佳實務調整，實現了每月高達 84% 的顯著費用節省。

三、彈性架構與效能突破：結合讀庫 Auto Scaling、自訂讀取端點與 Aurora Serverless，將每日高負荷任務的執行時間從 15 小時縮短至 2 小時。不僅實現了業務分離，更確保一般性任務的執行不受影響，發揮極佳的彈性價值。

在這場演講裡將會提到在金融相關產業使用 Debezium, Apache Kafka, Amazon Data Firehose, Amazon S3 Tables 和 Apache Iceberg 建立一個資料串流平台遇到了甚麼特別的問題以及如何解決

當 AI 都開始會自己拉架構、寫 CDK，甚至連 Code Review 都批得比你老闆還狠，身為 AWS 架構師團隊，到底會被 AI 賦能還是直接被「賦歸」？當 AI Agent 成為你的實習生甚至競業時，架構師該如何把 AI 當作專屬的黑魔法工具，守住飯碗的同時也讓生活更加愜意。

過去，我們習慣把生成式 AI 當成坐在旁邊的「副駕駛」：你下指令、AI 回答，你檢查結果，再決定下一步。當 AI 能保留工作脈絡、使用工具、分工處理任務，甚至在你離開座位後繼續工作，問題也跟著有了變化：哪些工作可以交給它？哪些決定，仍然得由你拍板？

這場分享以 Kiro Crew 為主軸，看看 AI 如何從隨問隨答的助理，成為能接續工作的隊友。我們會談長時間任務、記憶、子代理（subagents）與排程，也會討論權限與人工核准：它能讀什麼、改什麼，什麼時候必須先問過你。

讓它加班，別讓它加戲。除了讓 AI 多做一點，也可以一起思考哪些事，不能直接交給它。

AI agents are being adopted across industries, but building and operating them efficiently remains a major challenge. This session shares how an agentic AI assistant was first built with AWS Lambda, then evolved to Amazon Bedrock AgentCore as scalability, orchestration, and operational requirements grew. Through an honest production case study, attendees will see the architecture in action and learn practical patterns, migration decisions, development workflows, and operational lessons for building a fully serverless agentic AI platform.

Learn the importance of data and right storage platform for your AI workloads. The session will focus on the importance of data accessibility and mobility when building AI applications or agents.

Serverless gives us real power, but only if we accept its limits. They cover execution time, resources, package size, concurrency, and API rates. In return we get managed services we did not have to build. Those limits also match what a good design should do anyway. Misread them, and the same design starts to cost us. We once said "fully serverless" and meant different systems. The word was shared; the constraints behind it were not. AI coding tools lowered the barrier to building on serverless. They still need the experience we never wrote down. This talk is that record.

You will learn to ask whether you really want a fifteen-minute function, and to write down the limits you accepted.