Building a Personal AI Agent with Strands Agents × Amazon Bedrock AgentCore

Transcript

1. Building a Personal AI Agent with Strands Agents × Amazon Bedrock AgentCore Building a Personal AI Agent with Strands Agents × Amazon Bedrock AgentCore 2026/4/10
2. Please call me “yoko”! Name： yoko / Naoki Yokomachi / 横町 直樹 X：@_cityside Role： Software Engineer AWS Community Builders 2026 AI Engineering Category Affiliation： KDDI Agile Development Center, Inc. Zenn： https://zenn.dev/yokomachi Dev.to： https://dev.to/yokomachi
3. My First External Talk My First External Talk My First External Talk Strands Agents Meetup in July last year
4. What I’ll Talk About Today What I’ll Talk About Today What I’ll Talk About Today A Guide to Building a Personal AI Agent with Strands Agents and Amazon Bedrock AgentCore I gave an LT at the same venue last month↓ Focusing on the agent component of personal AI, introducing a real-world build
5. Agenda Demo Architecture Agent Implementation Syncing 3D Model Control with Agent Responses Too Many Tools Are Eating Up Tokens Code Execution Leveraging Controllability and Flexibility with Skills and Code Interpreter Features In Development (Planned)
6. Demo Demo Demo
7. Architecture Architecture Architecture
8. 　 Fro n te n d o n Ve rce l 　 　 ・ A I T u b e r K i t ： 3 D M o d e l C o n tr o l 　 　 ・ P o r c u p in e ： W a k e W o rd D e te c ti o n 　 　 ・ W e b S p e e c h A P I ： V o ic e I n p u t 　 　 ・ A i v i s S p e e c h ： L o c a l T T S 　 　 ・ O t h e r U t i l it y F e a t u r e s
9. 　 A ge n t o n B e d ro c k A ge n t C o re 　 　 ・ A g e n t : 　 　 　 - M a i n A g e n t 　 　 　 - S u b - a g e n t s s p li t b y d o m ai n 　 　 　 　 （ A g e n t s a s T o o l s） 　 　 　 - P i p e l i n e ag e n t s f o r s c h e du l e d e xe c u t i o n 　 　 　 　 （ n e w s n o t if i c a t i o n s a n d T w it t e r po s t s） 　 　 ・ E x t e r n a l I n t e g r a t i o n T o o l s 　 　 ・ S h o r t - t e r m M e m o r y / L o n g - t e rm M e m o r y 　 　 ・ S e c r e t M a n ag e m e n t 　 　 ・ S a n d b o x e d C o d e E x e c u t i on E n vi r o n m en t 　 　 ・ T u r n i n g L a m bd a F u n c t io n s in t o T oo ls
10. 　 B a c ke n d o n A WS 　 　 ・ L a m b d a ： I m p le m e n tati o n o f age n t to o l s a n d exe c u t i o n o f p i p el i n e a ge n ts 　 　 ・ E v e n t B r i d g e ： S c h e d u l e d e xe c u t i o n 　 　 ・ S N S ： N e w s n o t i f ic a t io n s 　 　 ・ D y a n a m o D B ： S t o r ag e fo r t a sk s , di a ri es , e tc . 　 　 ・ P o l l y ： B a c k e n d T T S 　 　 ・ C o g n i t o ： A p p - A g e n t C o re a u t h en t i c a ti o n 　 　 　 　 　 　 　 　 A p p - A P I G a t e wa y a u t he n t i c at i o n
11. Tool Integration
    Computing /
    Runtime
    Simple Implementation Same categor y as LangGraph, OpenAI Agent SDK, Mastra, etc. I ts sel l i ng
    poi nt
    at
    l aunch
    was
    that
    you
    can
    run
    an
    agent
    i n
    j ust
    3
    l i nes
    of
    code. Agentic Loop Repeats model
    cal l
    →
    tool
    i nvocati on
    →
    model
    cal l
    to
    generate
    responses Runtime
    Agent execution
    environment
    AWS Agent Framework Strands Agents Serverless Agent Platform Amazon Bedrock
    AgentCore Agent
    Implementation Agent
    Implementation Agent Implementation Code Interpreter‌ Code execution
    Browser
    Web browsing
    State Management
    Security &
    Governance
    Operation &
    Quality
    Management
    Gateway
    Turn APIs
    /
    Lambda
    /
    MCP
    into
    agent
    tools
    Memory
    Short-term &
    Long-term
    memory Identity
    Agent authentication
    Policy
    Guardrails for
    tool
    execution
    Observability
    Tracing &
    Monitoring
    Evaluations
    Continuous quality
    assessment
    of
    agents
    
12. [happy] Hel l o! How are you feel i ng today? Syncing 3D Model Control with Agent Responses Syncing 3D Model Control with Agent Responses Syncing 3D Model Control with Agent Responses Detect expressi on tags and appl y [bow] Understood! Let me i ntroduce mysel f! Detect moti on tags and bow [happy][wave_hand] Thank you very much! Pl ay expressi on and moti on si mul taneousl y As an interactive element unique to 3D models, play expressions and motions matched to responses 1.Instruct via prompt to tag responses 2. Frontend detects tags and plays corresponding expressions and motions ※Adjusted to avoid conflicts with idle motion and blinking during normal state (this part was painstakingly tuned manually since it's hard to delegate even to coding agents)
13. Motion Feature Also Contributed Back to AITuber Kit Motion Feature Also Contributed Back to AITuber Kit Motion Feature Also Contributed Back to AITuber Kit
14. Anthoropic Claude Haiku 4.5 on Bedrock $ 1.00 / 1M input tokens $ 5.00 / 1M output tokens Grok 4.1 fast on OpenRouter $ 0.16 / 1M input tokens $ 0.50 / 1M output tokens Too Many Tools Are Eating Up Tokens Too Many Tools Are Eating Up Tokens Too Many Tools Are Eating Up Tokens After continuously adding tools, ended up with 28 tools Before cost reduction: exceeded budget of $120/month Fixed token count increased 1.Create tool agents per domain 2.Change the model E s t i m a t e d ~ $ 6 0 r e d u c t i o n
15. ‌ Retrieve AWS
    costs
    with
    the
    cost
    management
    agent
    Visualize costs retrieved by the cost management agent using Code Interpreter
    Save the
    generated
    graph
    to
    S3
    and
    return
    its
    Presigned
    URL ‌ Detects the
    Presigned
    URL
    in
    the
    agent's
    response,
    fetches
    it
    from
    S3,
    and
    displays
    it
    Workflow order
    and
    sample
    code
    for
    graph
    generation
    are
    specified
    in
    Skills Code
    Execution
    Leveraging
    Controllability
    and Code
    Execution
    Leveraging
    Controllability
    and Flexibility
    with
    Skills
    and
    Code
    Interpreter Flexibility
    with
    Skills
    and
    Code
    Interpreter Code Execution Leveraging Controllability and Flexibility with
    Skills
    and
    Code
    Interpreter h t t p s : / / s t r a n d s a g e n t s . c o m / d o c s / u s e r - g u i d e / c o n c e p t s / p l u g i n s / s k i l l s / A s o f M a r c h 2 0 2 6 , S t r a n d s A g e n t s a l s o s u p p o r t s S k i l l s — e x t e r n a l p r o m p t s t h a t d o n ' t b l o a t t h e c o n t e x t . B u i l t a
    s m a l l
    w o r k f l o w
    w h i l e
    t e s t i n g
    C o d e
    I n t e r p r e t e r
    f e a t u r e s . D e t a i l s a r e
    p o s t e d
    o n
    Z e n n
    a n d
    D e v . t o ①aws-cost Skill ②Frontend Control " D r a w a
    r e d
    l i n e
    a t
    $ 5 " " S h o w m e
    t h e
    b r e a k d o w n
    b y
    s e r v i c e " Tool usage and order are preconfigured The agent decides which data to use and how to visualize
16. External Integrations: ☑ Web Search ☑ Google Calendar & Gmail Integration ☑ Notion Integration ☑ Twitter Integration ☐ Google Fit Integration ☐ Reference my own GitHub repos & suggest improvements to myself Extending Embodiment: ☑ Ears (wake word detection & voice input) ☑ Eyes (camera access) ☑ TTS (Aivis Speech, Polly) ☐ Camera angle and other controls ☐ Robotic arm control Agent Features: ☑ Task List (agent-accessible) ☑ Diary (agent-accessible) ☑ AWS Cost Visualization ☑ Briefing Feature Scheduled Tasks: ☑ News notification once/day ☑ X posts 4 times/day LLM: ☑ Amazon Bedrock Anthropic Claude models ☑ OpenRouter OpenAI-compatible models Utility Features: ☑ Pomodoro Timer ☑ Model Switching ☑ Dark Mode / UI Animation Toggle Features In Development (Planned) Features In Development (Planned) Features In Development (Planned) 🤔 Turns out there are quite a few features that are actually cumbersome to customize with just Claude Code → A personal agent you can freely build and break is great for both experimentation and learning M e l a s t m o n t h : " I t h i n k m o s t o f t h i s c o u l d p r o b a b l y b e d o n e w i t h C l a u d e C o d e , l o l "
17. Thank you!