sample-strands-agent-with-agentcore

# Strands Agent Chatbot with Amazon Bedrock AgentCore An end-to-end reference architecture for building **agentic workflows** using **Strands Agents** and **Amazon Bedrock AgentCore**. This repository demonstrates how to design and deploy a multi-agent chatbot that combines tool execution, memory, browser automation, and agent-to-agent collaboration. It is intended as a realistic, extensible sample for teams exploring advanced agent architectures on AWS. > [!NOTE] > **Updated Bedrock models.** The model picker includes the OpenAI **GPT-6 family (Astra, Sol, and Luna)** through the Bedrock Responses API (Sol/Luna via Mantle in `us-east-1`; Astra via Runtime) and **xAI Grok 4.6** through Bedrock Runtime Converse. **Claude Opus 5.5** uses Mantle Messages in `us-east-1`; **Google Gemma 4** remains available through Mantle in `us-east-2`. --- ## What this repository demonstrates - Multi-agent orchestration with **Strands Agents** - Integration with **Amazon Bedrock AgentCore** - Tool-enabled agents (search, finance, weather, browser, code interpreter) - Autonomous browser, documentation, and analysis workflows - Modular architecture adaptable to real customer use cases - Infrastructure-as-Code (Terraform) for repeatable deployment If you are building **agentic AI applications on AWS** and want a concrete, end-to-end example, this repository is designed to be read, run, and extended. **Quick links:** [Architecture](#architecture-overview) · [Key Features](#key-features) · [Quick Start](#quick-start) --- ## Architecture Overview This sample combines **Strands Agent orchestration** with **Amazon Bedrock AgentCore services**: | Component | Role | |-----------|------| | **Strands Agents** | Multi-turn reasoning and tool orchestration | | **AgentCore Runtime** | Managed, containerized agent execution | | **AgentCore Memory** | Persistent conversation state and summarization | | **AgentCore Gateway** | MCP-based tool integration with JWT authentication | | **AgentCore Code Interpreter** | Secure Python execution for analysis and document generation | | **AgentCore Browser** | Headless browser automation with live view | | **AgentCore Identity** | End-user authentication and 3LO OAuth delegation | | **AgentCore Registry** | Central catalog for agent skills, MCP servers, and A2A agents | | **AgentCore Observability** | Trace collection and agent execution monitoring | | **Amazon Nova Act** | Visual reasoning model for browser automation | <img src="docs/images/architecture-overview.png" alt="Architecture Overview" width="1200"> --- ## Key Features - Strands-based agent orchestration - Amazon Bedrock AgentCore Runtime, Gateway, and Memory - MCP Gateway tools (Google, Wikipedia, ArXiv, Yahoo Finance, Tavily, Open-Meteo) - Agent-to-Agent (A2A) collaboration — including remote **Claude Agent SDK (Claude Code)** for agentic coding tasks - Built-in Code Interpreter for charts and documents - Multimodal input and output (vision, charts, documents, screenshots) - Real-time voice interaction with Amazon Nova Sonic 2 - Wide model selection — Bedrock Runtime Converse models, GPT-6 via Bedrock Responses, and Gemma 4 via Bedrock Mantle --- ## Skill System (Progressive Disclosure) Tools are organized into **skills** — grouped units with SKILL.md instructions that the agent loads on demand. Instead of injecting all tool documentation into every prompt, the agent activates only the skills it needs: 1. **L1 Catalog** — Skill names and one-line descriptions (always in system prompt) 2. **L2 Instructions** — Full SKILL.md loaded via `skill_dispatcher` when activated 3. **L3 Execution** — Tool calls via `skill_executor` This keeps prompt size small while giving the agent access to detailed instructions when needed. Design notes: - https://medium.com/towards-artificial-intelligence/agent-skills-part-2-bridging-skills-with-production-tool-ecosystems-422e4a63fcad ``` skills/ ├── visual-design/ # Charts, posters, infographics (Code Interpreter) ├── code-interpreter/ # General code execution ├── browser-automation/ # Nova Act browser tools ├── word-documents/ # Word document generation ├── excel-spreadsheets/ # Excel spreadsheet generation ├── powerpoint-presentations/ # PowerPoint generation ├── gmail/ # Gmail read/search/delete (3LO OAuth) ├── google-calendar/ # Calendar events (3LO OAuth) ├── notion/ # Notion pages and databases (3LO OAuth) ├── github/ # GitHub repos, issues, PRs, code (3LO OAuth) └── ... # 20+ skills (web search, finance, maps, weather, and more) ``` --- ## Multi-Protocol Tool Architecture | Tool Category | Protocol | Examples | Authentication | |--------------|----------|----------|----------------| | Local Tools | Direct Python | Web Search, URL Fetcher, Visualization | None | | Built-in Tools | AWS SDK / WebSocket | Code Interpreter, Browser (Nova Act) | IAM | | Gateway Tools | MCP | Google Search, Maps, Wikipedia, ArXiv, Finance | SigV4 | | Private API Tools | MCP (3LO OAuth) | Gmail, Google Calendar, Notion, GitHub | OAuth 2.0 | | A2A Tools | A2A | Research Agent, Browser-Use Agent | JWT | Total: **100+ tools across 20 tool groups** See [docs/guides/TOOLS.md](docs/guides/TOOLS.md) for full details. --- ## Voice Mode Real-time voice interaction using **Amazon Nova Sonic 2**, **Strands BidiAgent**, and **AgentCore Runtime WebSocket**. - Seamless switching between voice and text within a single session - Shared conversation history across both modes - Full tool execution support during voice conversations --- ## Memory Architecture and Long-Context Management The system uses **AgentCore Memory** with: - Short-term session memory - Long-term summarized memory Long conversations are compacted using a context summarization strategy to retain key information while controlling token growth. Design notes: - https://medium.com/@revoir07/long-context-compaction-for-ai-agents-part-1-design-principles-2bf4a5748154 --- ## Token Optimization via Prompt Caching Prompt caching reduces input token usage by reusing system prompts, stable instruction blocks, and repeated conversation context across agent loop iterations. This project originally implemented caching via custom Strands hooks. The approach has since been upstreamed into the Strands SDK as a built-in feature ([strands-agents/sdk-python#1438](https://github.com/strands-agents/sdk-python/pull/1438)): ```python from strands.models import BedrockModel, CacheConfig model = BedrockModel( model_id="us.anthropic.claude-sonnet-5", cache_config=CacheConfig(strategy="auto") ) ``` Design notes: - https://medium.com/@revoir07/agent-loop-caching-the-missing-optimization-for-agent-workflows-230cc530eb72 --- ## Multi-Agent Architecture Agent-to-Agent communication is handled via the **A2A protocol**, allowing the supervisor agent to delegate tasks to specialized worker agents such as a deep research agent. Multiple agents collaborating in sequence — each handling its own role — are coordinated using the **Swarm** pattern. This architecture also includes a **Claude Agent SDK (Claude Code)** instance deployed as a remote A2A agent on **AgentCore Runtime**. The supervisor delegates agentic coding tasks — multi-file implementation, refactoring, test suites — to this agent over A2A, with an S3-backed workspace that persists files across sessions. Design notes: - https://medium.com/@revoir07/extend-your-chatbot-with-deep-research-using-a2a-ba4de3ed23e9 --- ## Use Cases - Financial research agents - Technical research assistants using multi-agent patterns - Autonomous web automation agents - Memory-backed conversational assistants - Hybrid research workflows using MCP, A2A, and AWS SDK tools --- ## Quick Start ### Prerequisites - AWS account with Bedrock access - AWS CLI configured - Docker installed - Node.js 18+ and Python 3.13+ --- ### Local Development ```bash git clone https://github.com/aws-samples/sample-strands-agent-with-agentcore.git cd sample-strands-agent-with-agentcore cd chatbot-app ./setup.sh # Create chatbot-app/.env with deployment outputs (Cognito IDs, API URL, etc.) # — typically populated from terraform output after `infra/scripts/deploy.sh apply`. ./start.sh ``` Frontend will be available at http://localhost:3000. --- ### Cloud Deployment ```bash # Configure cp infra/environments/dev/terraform.tfvars.example infra/environments/dev/terraform.tfvars # Edit terraform.tfvars with your settings # Deploy all ./infra/scripts/deploy.sh apply # Deploy specific modules ./infra/scripts/deploy.sh apply -target=module.chat ./infra/scripts/deploy.sh apply -target=module.runtime_orchestrator ``` See [DEPLOYMENT.md](DEPLOYMENT.md) for full instructions. ### Project Structure ``` sample-strands-agent-chatbot/ ├── chatbot-app/ │ ├── frontend/ # Next.js UI + BFF │ └── agentcore/ # Python backend (Strands Agent) ├── agentcore/ # AgentCore workloads (built/deployed by Terraform) │ ├── gateway-tools/ # Lambda MCP tools (arxiv, weather, tavily, ...) │ ├── a2a-agents/ # A2A runtimes (code-agent, research-agent) │ └── mcp-runtime/ # MCP 3LO runtime (Gmail, GitHub, Notion, ...) └── infra/ # Terraform infrastructure ├── modules/ # Reusable modules (auth, runtime, gateway, chat, ...) ├── environments/dev/ # Environment configuration └── scripts/ # Deploy orchestrator ``` ## Documentation - DEPLOYMENT.md - docs/guides/GOOGLE_OAUTH_SETUP.md - docs/guides/GITHUB_OAUTH_SETUP.md - docs/guides/THREE_LEGGED_OAUTH_FLOW.md - docs/guides/TROUBLESHOOTING.md --- ## License MIT License. See LICENSE for details.