AI Agent that supports ”YOU” / introducing-bedrock-engineer-en
Universal AI agent building apps using Amazon Bedrock, capable of customize to create/edit files, execute commands, search the web, use knowledge base, use multi-agents, generative images and more.
rights reserved. Amazon Confidential and Trademark. Daisuke Awaji Amazon Web Services Japan Solutions Architect 2 @gee0awa Serverless, Generative AI, Frontend ❤ @gee0awa 👆 Check the resources from here
rights reserved. Amazon Confidential and Trademark. • Software development isn't just about coding... • LLMs don't understand how to use new libraries that they haven't been trained on • Development agents don't understand software architecture, coding conventions, or the development team's vision/intentions 4 Perspective on Development Agents @gee0awa 👆 Check the resources from here
rights reserved. Amazon Confidential and Trademark. • Software development isn't just about coding... It would be great if it could understand existing source code, create diagrams, and explain them! • LLMs don't understand how to use new libraries they haven't been trained on It would be great if they could implement while searching for the latest information! • Development agents don't understand software architecture, coding conventions, or the development team's vision it would be great if we could customize how development agents work! 5 Thoughts about Development Agents Generative AI can be used with serverless I created it using Amazon Bedrock Serverless Buddy @gee0awa 👆 Check the resources from here
rights reserved. Amazon Confidential and Trademark. Bedrock Engineer 6 https://github.com/aws-samples/bedrock-engineer An AI assistant that supports development using Amazon Bedrock
rights reserved. Amazon Confidential and Trademark. 8 Installation Method (when building from source) $ npm install $ npm run build:mac npm run build:win npm run build:linux または または Clone the repository and execute two commands A native app will be built on your local machine! ① Build it ② Open the app Double-click to open Please use the correct command according to your OS: Mac, Windows, and Linux Getting Started
rights reserved. Amazon Confidential and Trademark. You can start using it right away after setting up your AWS access key and secret access key from the settings screen 9 Initial Configuration Open settings screen Set up access key and secret access key Getting Started
rights reserved. Amazon Confidential and Trademark. In the AWS Management Console, select Bedrock model access, then request access to your desired models through “Modify model access” 10 Enable Bedrock Model Access
rights reserved. Amazon Confidential and Trademark. 12 AI Agent An AI system or program that understands its own environment and accomplishes tasks toward its goals by utilizing various tools https://www.anthropic.com/research/building-effective-agents
rights reserved. Amazon Confidential and Trademark. 13 General-purpose Development Agent • Reads source code, creates folders, and implements code. • Understands the source code structure, creates diagrams in Mermaid.js format, and supports developers. • Uses web search and RAG (Retrieval-Augmented Generation) search to consider optimal implementation plans according to the implementation context. ▶ ◀ Read Write
rights reserved. Amazon Confidential and Trademark. 14 Agent Selection Selection of tools to use Selection of working folders System prompt verification Agent customization Chat history
rights reserved. Amazon Confidential and Trademark. 15 (Development Example) Diagram and Explain Software Structure ① Read the software's directory structure/files ② Understand the structure, create diagrams and explain (This diagram expresses the sequence of agent chat) ③ さらに実装をしたり、設計書を書いたり、、、
rights reserved. Amazon Confidential and Trademark. 16 Available Tools Agent Tools User File reading File writing File moving, copying Folder creation URL-specified search Searching using Tavily Search Image generation Amazon Nova, Titan Stable Diffusion series Execute arbitrary commands Amazon Bedrock Knowledge Base Agents Custom tools can be implemented and integrated Python API Database Equipped with common tools for AI agent development and usage, can connect to databases and APIs through arbitrary commands and MCP MCP Server Model Context Protocol
rights reserved. Amazon Confidential and Trademark. 17 ReAct Agent User Chat Interface Prompt Tools Interprets and resolves the purpose of user queries. Available tools：[…] Reasoning & Action history: [ …, …, …, ] Repeat until the solution is found. LLM Was the user's objective solved? Analise Reasoning User Query No Yes Answer 1 2 3 4 5 N times Bedrock Engineer's AgentChat proceeds with processing using ReAct (Reasoning & Acting) steps Add to history ReAct (Reasoning & Actinng)
rights reserved. Amazon Confidential and Trademark. 19 Implementation Examples of Bedrock Converse API Tool Use toolSpec definition（Sent to Bedrock） Implementation of the Tool (Function) to be executed Bedrock Engineer extends various functionalities by using ToolUse in its internal processing.
rights reserved. Amazon Confidential and Trademark. 20 Bedrock Converse API The AI agent's operation is achieved by constructing messages array containing LLM Request/Response and ToolUse Input/Output, and sending it to Amazon Bedrock's Converse API. [ { "role": "user", "content": [{"text": " What's the weather in Sumida Ward, Tokyo?"}], }, { "role": "assistant", "content": [ { "toolUse": { "input": {"city": ”Sumida Ward", "prefecture": ”Tokyo"}, "name": "get_weather", "toolUseId": "tooluse_UwHeZGCnSQusfLrwCp9CcQ", } }, ], }, { "role": "user", "content": [ { "toolResult": { "content": [{"text": "In Sumida Ward, Tokyo, the weather is sunny, with a temperature high of 22 degrees."}] "toolUseId": "tooluse_UwHeZGCnSQusfLrwCp9CcQ", } } ], }, { "role": "assistant", "content": [ { "text": "I understand. I hear that in Sumida Ward, Tokyo, the weather is sunny with a high temperature of 22 degrees. With the sunny weather and relatively warm temperature, it seems like it will be a comfortable day. When going out, light clothing is recommended." } ], }, ] 0 1 2 3 ・・・ Message from user ・LLM response ・ToolUse input ・Function execution results specified by ToolUse ・ToolUse output Message from LLM
rights reserved. Amazon Confidential and Trademark. 21 What happens when the message array gets longer? When a conversation becomes too long to fit within the Context Length (for example, 200K tokens), many development agents implement a solution where they delete (or summarize) unnecessary elements from the message array before sending to the LLM. When executing long-running tasks, there's a risk of falling into a loop where past execution results are forgotten and tasks are re-executed. 0 1 2 3 ・・・ Message from user ・LLM response ・ToolUse input ・Function execution results specified by ToolUse ・ToolUse output Message from LLM 400 401 402 403 ・・・ 100 101 102 103 Keep only the initial instructions (to be sent to the LLM) Delete or summarize unnecessary steps (exclude from content to be sent to the LLM) Keep the context necessary for the currently ongoing task (to be included in content to be sent to the LLM)
rights reserved. Amazon Confidential and Trademark. 22 Guidelines for Context Length In Bedrock Engineer, the context length (length of the message array to be sent) can be configured. It depends on the information volume and token count contained in each message element. Assuming an agent that reads one file per message, the list below shows the various tasks and their corresponding required array lengths. This length Array length Task Type Task Examples 30 Very small-scale tasks Changes that can be completed with 1-2 files, like “Modify the processing of ~ in file ~ to do ~” 60 Small-scale tasks Cause analysis with human assistance, such as “Investigate the cause of [specific] problem. Focus your search in files [X] and [Y] 90 Functional changes that would fit in a single commit “After carefully investigating the current implementation, implement [specific] functionality” 120 Excessive use of tokens The work being executed at once is too large, so tasks should be broken down into smaller ones Implement [specific] functionality. Then, without waiting for my confirmation, implement cross- platform i18n support. Finally, for completeness check... (etc.)
rights reserved. Amazon Confidential and Trademark. By utilizing Tavily Search and URL-specified searches, we can create an agent that performs performs tasks while acquiring necessary knowledge based on the context. 23 Web Grounding Example showing how to supplement knowledge from the web during development Autonomously modify queries to search the web until the required information is gathered 1） How do you implement routing in React? 3） What are the implementation methods other than react-router? 2） What are the implementation methods other than react-router? Searching with specified URLs Searching using Tavily Search Tool Capabilities
rights reserved. Amazon Confidential and Trademark. 24 Web Grounding ｜Use Case • Research and implement using latest documentation for CloudFormation properties and AWS CDK Constructs • Examine error logs from software development process, determine corrective actions, and modify code accordingly Important Considerations • Using resources such as the Knowledge Base or AWS Document MCP server (covered later) might provide better search capabilities when investigations need to be confined to specific documents.
rights reserved. Amazon Confidential and Trademark. Unlike traditional RAG which required manual query refinement, the Agent-based approach automatically optimizes and executes queries efficiently. 25 Agentic-RAG Execute multiple optimized queries against the Knowledge Base “Admin Interface: tables, forms, and data visualization” Tool Capabilities Amazon Bedrock Knowledge Base “Layout Components: ResponsiveLayout AppLayout Header” （First query） （Second query） Create an e-commerce management console using AWS design patterns Code samples need to be prepared in advance
rights reserved. Amazon Confidential and Trademark. Create the basic structure and layout for an e-commerce website specializing in potted plants, following the CloudScape Design System. Let's build it using React, TypeScript, and Vite. The main requirements are as follows: <Conditions> •The layout should be similar to Amazon.com. •Set the e-commerce website's name as 'Green Village’. •Use a green-based color scheme. •Add a section featuring plant products displayed as product cards. •Add functionality to add items to a shopping cart. •Add functionality to review current shopping cart contents and compute the total purchase amount. </Conditions> 26 (Example of Agentic-RAG application) Web Application with Design Conformance Prompt Generated Application By providing reference source code in the Knowledge Base, generated applications will conform to your design guidelines. Agent (Bedrock Engineer) Knowledge Base （First query） Knowledge Base （Second query） Tool Capabilities
rights reserved. Amazon Confidential and Trademark. 27 Agentic-RAG ｜Use Case • Developing a website with UI that conforms with design guidelines • Developing code according to company coding standards
rights reserved. Amazon Confidential and Trademark. Using Bedrock Engineer as an orchestrator, multiple Sub-Agents can be executed. Sub-Agents can be used via Bedrock Agent or executed through CLI. 28 Multi-Agent Orchestration Sub-Agent (Bedrock Agent) Agent (Bedrock Engineer) Sub-Agent (via CLI) Example: Using Bedrock Agent's Code Interpreter for CSV sales data visualization Tool Capabilities
rights reserved. Amazon Confidential and Trademark. ② Tool can now be used 29 Connecting to MCP Servers Connecting to MCP servers allows access to external resources and tools. ① Input MCP Server settings Tool Capabilities
rights reserved. Amazon Confidential and Trademark. 30 MCP Server Examples Provided by awslabs AWS Documentation MCP •Read Documentation: Fetch & convert AWS doc pages to markdown format •Search Documentation: Search AWS documentation using the official search API •Recommendations: Get content recommendations for AWS documentation pages Knowledge Bases Retrieval MCP Server •Discover Knowledge Bases: Search for available knowledge bases and data sources •Natural Language Query: Query knowledge bases using natural language •Data Source Filtering: Focused search on specific data sources •Search Result Reranking: Improve relevance with Amazon Bedrock's reranking feature AWS Cost Analysis MCP Server •Analyze and Visualize AWS Cost: Display cost breakdown by service, region, and tier •Query Cost Data with Natural Language: Ask about your AWS costs in plain English •Cost Optimization Suggestions: Create cost reports & provide optimization proposals AWS CDK MCP Server •CDK General Guidance: Provide implementation patterns using AWS Solutions Constructs •CDK Nag Integration: Apply and explain security and compliance rules •AWS Solutions Constructs Support: Pattern search for common architectural needs •Generative AI CDK Constructs: Search for specialized constructs for AI/ML workloads And more…
rights reserved. Amazon Confidential and Trademark. 31 MCP Server Examples｜Use Cases • Inspect an AWS environment and draw the configuration diagram (e.g., draw.io format) • Analyze application error logs for root cause analysis • Research GitHub issues and propose fixes • Integration with project management tools like Jira to manage backlogs • Automate browser testing • Connect to application databases And more…
rights reserved. Amazon Confidential and Trademark. • Security MCP servers might include malicious code/network activity. This requires careful screening and adherence to proper usage guidelines. • Keep tools simple (number and naming of input variables). Just like humans, complex tools are difficult to master. Design with clear tool names, variable names, and manageable input values. • Keep tool execution results small. When there are large results, it may not fit within the context window. Even if it does, there is increased risks of Forgetting and Model Drift. For web searches, APIs, or databases, aim to design queries that maintain small result sets. 32 Points to consider when customizing tools
rights reserved. Amazon Confidential and Trademark. Rather than mandating extended thinking before analyzing code, let the AI agent decide when to engage in thinking during code analysis 33 “Think” Tool Tool Capabilities https://www.anthropic.com/engineering/claude-think-tool ToolSpec Example T-Bench performance evaluation results
rights reserved. Amazon Confidential and Trademark. • It's important that development agents can expand their capabilities through web searches, Agentic-RAG, and collaboration with other agents. • Development tools like Cline and Cursor can extend their functionality by connecting to MCP servers. • Consider if it's possible to create applications that can reference internal core modules and design system code • However, it's necessary to have controllable mechanisms such as guardrails to ensure security when executing tools. 34 What to Expect from Development Agents (Tools Section)
rights reserved. Amazon Confidential and Trademark. 36 Customizing the Agent You can create development agents specialized in specific fields, not just general-purpose development agents. You can make development agents that follow the development team's common conventions and standards, and flexibly control their behavior. For example, it's possible to have AI agents execute evaluation processes, such as automatically running 'npm run lint' after code generation to autonomously verify type errors.
rights reserved. Amazon Confidential and Trademark. 37 Customizing the Agent ① Insert the agent's name and description ② Click the System Prompt auto-generation button ③ The system will produce a prompt automatically, incorporating tool specifications and contextual considerations.
rights reserved. Amazon Confidential and Trademark. 38 Customizing the Agent (Example) Infrastructure as Code Expert You are an AI assistant specializing in AWS infrastructure. As an expert in CloudFormation and AWS CDK, you have the following capabilities: Infrastructure as Code (IaC) Design and Implementation: - Creation and optimization of AWS CloudFormation templates - Infrastructure definition using AWS CDK - Architecture design based on best practices - Security and compliance considerations - Cost optimization proposals Technical Expertise: - Creation of CloudFormation templates in YAML/JSON format - AWS CDK code development using TypeScript/Python/Java - Deep knowledge of AWS services and how to combine them - Design of multi-account/multi-region configurations - Integration of IaC into CI/CD pipelines Working Guidelines: - The project root directory is {{projectPath}} - Use absolute paths when reading and writing files - Use the read_file tool when analyzing existing code - Implement new files and changes with the writeToFile tool Providing Best Practices: - Design modular and reusable components - Implement methods for environment separation (development/staging/production) - Standardization of tagging and resource naming conventions - Design appropriate IAM policies and roles - Design efficient stack structures When Creating Projects: - Propose appropriate project structure - Create necessary directories and files - Provide basic configuration file templates - Support deployment pipeline design Providing Latest Information: - Use tavilySearch for researching AWS new features and updates Mobile App Development Expert Under what circumstances, which tools will be used Agent’s Purpose Agent’s Approach ToolSpec You are an expert AI assistant in TypeScript, React Native, Expo, and mobile UI development. You have specialized knowledge and skills in the following areas: - Type-safe development using TypeScript - Cross-platform mobile application development with React Native - Utilize Expo framework and its best practices - Mobile UI design patterns and user experience - Performance optimization and debugging - App lifecycle management and deployment Key strengths include: - Mobile application architecture design - Implement TypeScript and React Native best practices - Effective utilization of Expo toolchain - Create responsive and attractive UI components - Application performance optimization - Ensuring cross-platform compatibility Utilize the following tools for development support: - Creation and management of project structure (createFolder, writeToFile) - Analysis and review of existing code (readFiles, listFiles) - Obtaining latest technical information (tavilySearch, fetchWebsite) - Visualization of UI/UX design (generateImage) - Information retrieval from knowledge base (retrieve) - Execution of necessary commands (executeCommand) - Working directory: {{projectPath}} Important Rules for Project Creation: - Always start by creating the project root folder - Enforce implementation of TypeScript type definitions - Emphasize component reusability and testability - Properly manage Expo configuration and platform-specific settings - Always consider application performance and user experience - Adherence to security best practices When Doing Code Review and Enhancements: - Verify TypeScript type safety Additional Rules
rights reserved. Amazon Confidential and Trademark. 39 Agent Use Cases Outside of Development By controlling the system prompt of agents, they can be applied to non-development tasks as well. Agents that behave like café staff or psychological counselors are some examples. Example: Café Staff Agent You are a kind-hearted café staff AI assistant. You strive to provide attentive and caring service while being considerate of customers Personality and Attitude: - Provide service with a friendly and pleasant attitude - Polite and courteous language usage - Actively understand customer needs and make appropriate suggestions - Show particular consideration to customers who need assistance - Always strive to serve with a smile Basic Service Protocol: - Begin with a greeting to the customer - Take orders courteously - Provide menu explanations and suggestions - Check for allergies and preferences - Repeat and confirm orders - Guide through payment process - Farewell message Menu Knowledge: - Able to explain in detail the characteristics of coffee, tea, and other drinks - Understanding of food/dessert menu items and their preparation methods - Accurately provide allergy information and ingredients - Guidance on seasonal menu items Special Accommodations: - Utmost care for customers with allergies - Consideration for customers with children - Multilingual support for foreign customers - Support for elderly customers and those with disabilities
rights reserved. Amazon Confidential and Trademark. 40 (Custom Agent Example) Kamishibai* Expert User's instruction: Please create a story with an elephant as the main character to convey the importance of tooth brushing to a 3-year- old. 1) Create directory structure and establish story outline 2) Generate the most appropriate images for each kamishibai slide Automatically create and execute image generation LLM prompt requests * Kamishibai is a traditional form of Japanese storytelling that uses illustrated cards and a narrator to present stories, often in a theatrical style.
rights reserved. Amazon Confidential and Trademark. 43 Tool Customization | Adding Custom Tools You can connect agents to APIs and databases through the command execution tool. 43 Agent Tools User File reading File writing File moving, copying Folder creation URL-specified search Searching using Tavily Search Image generation Amazon Nova, Titan Stable Diffusion series Custom Command Execution Retrieve Query to Amazon Bedrock Knowledge Base Amazon Bedrock Agents Execution Custom tools can be implemented and integrated Python API Database
rights reserved. Amazon Confidential and Trademark. 44 Tool Customization Example: 1) API Data Analysis An AI agent that provides policy decision-making insights using APIs to obtain data on medical care, demographics, and nursing care facilities Instructions: Utilize nursing care facilities, medical facilities, demographic, and regional data to assist in future policy decision-making for Tokyo Metropolitan area. Start by collecting data and investigate what issues exist. (Bedrock Engineer) Fetch and collect Tokyo Metropolitan data via API and analyze for issues.
rights reserved. Amazon Confidential and Trademark. 45 Tool Customization Example 2) AWS Resource Analysis By enabling AWS CLI execution, this enables visualization of resources running on AWS and log analysis. Example: Retrieval of list of AWS Lambdas
rights reserved. Amazon Confidential and Trademark. 46 Tool Customization Example 3) Browser Use Run the ‘browser-use' agent, which controls the browser, via the command line. Search Amazon and suggest products for Daisuke Awaji of AWS. (Prerequisites) Set up the browser control agent to enable browser-use execution via the executeCommand tool https://github.com/browser-use/browser-use Browser session showing “cooking gadgets” recommended for Daisuke Awaji Browser operation
rights reserved. Amazon Confidential and Trademark. Content hub for sharing agents/system prompts developed by users 48 Agent Directory You can select an agent and use it immediately
rights reserved. Amazon Confidential and Trademark. 49 Application Guardrails Agents autonomously read files and collect database information, but can return masked results and prevent transmission of sensitive data to LLMs. An example of masking an uploaded file From the Settings screen, specify the Guardrail identifier and version
rights reserved. Amazon Confidential and Trademark. Automatically apply guardrail by specifying in guardrailConfig 50 Converse API guardrailConfig Application Amazon Bedrock LLMs Converse API Guardrail is automatically applied ガードレールの適用方法１ (Similar settings are also possible with the InvokeModel API)
rights reserved. Amazon Confidential and Trademark. You can use Guardrail functions directly without going through LLM. This is effective in cases where you want to verify the execution results (ToolResult) of ToolUse before including them in the request, or when using an LLM that is not supported by Amazon Bedrock. 51 ApplyGuardrail API Application ApplyGuardrail In cases where you only want to use the guardrail ガードレールの適用方法２
rights reserved. Amazon Confidential and Trademark. 52 Prompt Caching A comparison with and without caching when instructed to 'Analyze the source code and explain its structure and function, including diagrams’ * Note that this is not a precise comparison due to slight variations in the total token count. Without Caching With Caching
rights reserved. Amazon Confidential and Trademark. 53 Cost A dashboard is provided that calculates costs based on the number of tokens used in sessions and cached tokens. These are reference values only, please refer to the AWS Management Console for accurate figures. https://aws.amazon.com/bedrock/pricing/
rights reserved. Amazon Confidential and Trademark. 54 Cost Estimates AWS Service Dimensions Cost [USD/month] Bedrock Input: Sonnet 3.7 400k tokens/session 120 Bedrock Output: Sonnet 3.7 20k tokens/session 30 https://aws.amazon.com/jp/bedrock/pricing/ Assuming 100 sessions per month and the monthly cost being proportional to the number of sessions. Expected use cases such as refactoring existing code and performing web searches.
rights reserved. Amazon Confidential and Trademark. By separating AI assistant operations into two phases - ”Plan" and ”Execute", users can have more granular control over it’s behavior. 56 1. Planning and execution Plan Execute The ~ folder contains the implementation of ◦◦ function. Add the △△ function to this feature. The requirements are as follows: ・XXXX ・YYYY Start by analyzing the source code to develop an implementation strategy, then request my feedback. Start the implementation according to that plan. During the development process, execute commands for type checking, static analysis, and other related tasks.
rights reserved. Amazon Confidential and Trademark. 57 2. Guidelines for Context Length In Bedrock Engineer, the context length (length of the message array to be sent) can be configured. It depends on the information volume and token count contained in each message element. Assuming an agent that reads one file per message, the list below shows the various tasks and their corresponding required array lengths. Array length Task Type Task Examples 30 Very small-scale tasks Changes that can be completed with 1-2 files, like “Modify the processing of ~ in file ~ to do ~” 60 Small-scale tasks Cause analysis with human assistance, such as “Investigate the cause of [specific] problem. Focus your search in files [X] and [Y] 90 Functional changes that would fit in a single commit “After carefully investigating the current implementation, implement [specific] functionality” 120 Excessive use of tokens The work being executed at once is too large, so tasks should be broken down into smaller ones Implement [specific] functionality. Then, without waiting for my confirmation, implement cross- platform i18n support. Finally, for completeness check... (etc.)
rights reserved. Amazon Confidential and Trademark. 3. System Prompt Structure Include the agent’s purpose, restrictions, preferences, and tool usage guidelines. This allows for more targeted control and tool use, while conserving tokens. A primary cause of “I've set up the MCP server, but it's not behaving as expected!” is often the absence or vagueness of these system prompt directives. Infrastructure as Code Expert You are an AI assistant specializing in AWS infrastructure. As an expert in CloudFormation and AWS CDK, you have the following capabilities: Infrastructure as Code (IaC) Design and Implementation: - Creation and optimization of AWS CloudFormation templates - Infrastructure definition using AWS CDK - Architecture design based on best practices - Security and compliance considerations - Cost optimization proposals Technical Expertise: - Creation of CloudFormation templates in YAML/JSON format - AWS CDK code development using TypeScript/Python/Java - Deep knowledge of AWS services and how to combine them - Design of multi-account/multi-region configurations - Integration of IaC into CI/CD pipelines Working Guidelines: - The project root directory is {{projectPath}} - Use absolute paths when reading and writing files - Use the read_file tool when analyzing existing code - Implement new files and changes with the writeToFile tool Providing Best Practices: - Design modular and reusable components - Implement methods for environment separation (development/staging/production) - Standardization of tagging and resource naming conventions - Design appropriate IAM policies and roles - Design efficient stack structures Mobile App Development Expert Under what circumstances, which tools will be used Agent’s Purpose Agent’s Approach ToolSpec You are an expert AI assistant in TypeScript, React Native, Expo, and mobile UI development. You have specialized knowledge and skills in the following areas: - Type-safe development using TypeScript - Cross-platform mobile application development with React Native - Utilize Expo framework and its best practices - Mobile UI design patterns and user experience - Performance optimization and debugging - App lifecycle management and deployment Key strengths include: - Mobile application architecture design - Implement TypeScript and React Native best practices - Effective utilization of Expo toolchain - Create responsive and attractive UI components - Application performance optimization - Ensuring cross-platform compatibility Utilize the following tools for development support: - Creation and management of project structure (createFolder, writeToFile) - Analysis and review of existing code (readFiles, listFiles) - Obtaining latest technical information (tavilySearch, fetchWebsite) - Visualization of UI/UX design (generateImage) - Information retrieval from knowledge base (retrieve) - Execution of necessary commands (executeCommand) - Working directory: {{projectPath}} Important Rules for Project Creation: - Always start by creating the project root folder
rights reserved. Amazon Confidential and Trademark. Bedrock Engineer Architecture 61 Web UI exposeInMainWorld fetch api.create main handlers (createFile, readFile…) api.listen(port) app.ready create.window Amazon Bedrock AWS SDK Web API IPC (Inter-Process Communication) Electron (Native app for macOS, Win, Linux) AWS Cloud main renderer preload Implemented as a native application that runs on Mac, Windows, and Linux using Electron For AWS cloud services, only Amazon Bedrock is used, with a serverless configuration and no AWS Lambda requirement. CDK Deployment is also not necessary.
rights reserved. Amazon Confidential and Trademark. Bedrock Engineer Architecture (if web based) 62 Web UI exposeInMainWorld fetch api.create main handlers (createFile, readFile…) api.listen(port) app.ready create.window Amazon Bedrock AWS SDK Web API IPC (Inter-Process Communication) Electron (Native app for macOS, Win, Linux) AWS Cloud Web UI fetch Amazon Bedrock Web API SPA (Single Page Application) AWS Cloud AWS Lambda Amazon CloudFront (with S3) Retrieval of static content main renderer preload Electron’s React-based UI allows for easy web conversion as long as no file operations (OS operations) are required. Use AWS Lambda for handling tool executions and LLM requests
rights reserved. Amazon Confidential and Trademark. 64 Website Generation Examples E-commerce Site for Indoor Plants Data Retrieval via API and Data Visualization Healthcare Blog Generates and previews React, Vue, and Svelte code in real-time. Enables website creation and feature enhancement through interactive sessions.
rights reserved. Amazon Confidential and Trademark. 65 Smart Recommendations for Website Generation Additional feature recommendations are proposed based on conversational context We've implemented a recommendation system by structuring GenAI outputs in JSON format. Smart Recommendations Amazon Bedrock GenAI foundation model Output in JSON array format that includes keys called "title" and "value". The following is an example. <example> [{title: ʼʼ, value: ʻʼ}] </example> [ { title: “Optimizing Imported Images”, value: ”For image loading...” }, { title: “Shopping cart enhancement”, value: “Shopping cart functionality...” } ] Prompt Output Incorporate JSON Schema for better results
rights reserved. Amazon Confidential and Trademark. 66 Connect to Knowledge Base Generate websites by referencing information and source code stored in Amazon Bedrock Knowledge Base Autonomously referencing design systems and core modules to produce code e-Commerce site for indoor plants that follows AWS Cloudscape Design System Connect with Knowledge Base
rights reserved. Amazon Confidential and Trademark. 67 Connect to the Web Autonomously researches web content and generates websites. Tasks such as 'Build my profile page' can be accomplished with ease. Create a profile page for AWS Solution Architect, Daisuke Awaji Enable Web Search
rights reserved. Amazon Confidential and Trademark. 72 Connect to Web Draw the system diagram after enabling web search and gathering needed information Enable Web Search
rights reserved. Amazon Confidential and Trademark. You can collect and visualize information about active AWS resources by enabling a Tool capable of running specific AWS CLI commands. This is an experimental feature - to use it, build from the source code in the diagram-generator branch 73 Connect to your AWS Environment Experim ental Example: Result of command 'Display the resources in the VPC in my AWS environment'
gawa
chrisshort
3n
keavy
michaelherold
jnunemaker
kevinliebholz
keathley
lara
mraible
marcduiker
orderedlist
sferik
