A comprehensive implementation of the Agent-to-Agent (A2A) protocol using specialized agents running on Amazon Bedrock AgentCore runtime, powering an intelligent educational platform for course material creation, student preparation, video analytics, and document management.
This repository walks you through setting up a multi-agent orchestration system where specialized agents collaborate via A2A to handle distinct educational workflows — Video Flow and Documents Flow — all coordinated by a central orchestrator and backed by AgentCore primitives (memory, observability, identity, and gateway).
The system is composed of two primary flows:
- A Video Agent (Strands SDK on AgentCore) exposes tools like
fetch_metadata,fetch_transcript,fetch_engagement,fetch_polls, andcompute_video_insights - Requests route through an AgentCore Gateway to a Fetch Video Data Lambda that calls the Company Video API
- Results are stored in a Bedrock Knowledge Base backed by an S3 bucket (
video-metadata) and a Video Metadata DynamoDB table - The agent has its own Memory and Observability via AgentCore
- A Documents Agent (Strands SDK on AgentCore) delegates to sub-agents:
- Document Context Agent — retrieves and manages educational documents via a Bedrock Knowledge Base with S3 Vector / Nova Multi-Modal Embeddings, reading from
e-ink-documentsS3 and the Company Documents API - Analytics Agent — tracks usage analytics, writing to Analytics Data (DynamoDB), with an analytics ingest Lambda pushing to the Company Analytics Data API and
ink-analytics-rawS3 - Session Manager — manages user sessions across interactions
- Document Context Agent — retrieves and manages educational documents via a Bedrock Knowledge Base with S3 Vector / Nova Multi-Modal Embeddings, reading from
- Each sub-component has its own Memory and Observability
- Users authenticate via Cognito and interact through an Amplify Frontend
- Requests hit API Gateway → Orchestrator Lambda → A2A Client which routes to the appropriate agent runtime
- An Admin Dashboard (also behind Cognito) provides proxy Lambdas for video/document analytics data and agent management
Note
Default Models
This solution uses the following AI models by default:
- Host Agent (Strands):
global.anthropic.claude-sonnet-4-5-20250929-v1:0(Amazon Bedrock) - Echo Ink Agent (Strands):
global.anthropic.claude-sonnet-4-5-20250929-v1:0(Amazon Bedrock) - Echo Prepare Agent (Strands):
global.anthropic.claude-sonnet-4-5-20250929-v1:0(Amazon Bedrock) - Monitoring Agent (Strands):
global.anthropic.claude-haiku-4-5-20251001-v1:0(Amazon Bedrock)
These models can be customized during deployment.
Agent-to-Agent (A2A)
Agent-to-Agent (A2A) is an open standard protocol that enables seamless communication and collaboration between AI agents across different platforms and implementations. The A2A protocol defines:
- Agent Discovery: Standardized agent cards that describe capabilities, skills, and communication endpoints
- Communication Format: JSON-RPC 2.0-based message format for reliable agent-to-agent communication
- Authentication: OAuth 2.0-based security model for secure inter-agent communication
- Interoperability: Platform-agnostic design allowing agents from different frameworks to collaborate
Learn more about the A2A protocol: A2A Specification
Amazon Bedrock AgentCore provides native support for the A2A protocol, enabling you to:
- Deploy A2A-compliant agents as runtime services with automatic endpoint management
- Secure authentication via AWS Cognito OAuth 2.0 integration
- Agent discovery through standardized agent card endpoints
- Scalable deployment leveraging AWS infrastructure for production workloads
- Built-in observability with CloudWatch integration and OpenTelemetry support
| Agent | Role | SDK | Runtime |
|---|---|---|---|
| Host / Orchestrator | Routes user requests to the correct downstream agent via A2A | Strands | AgentCore |
| Echo Ink | Creates educational documents (syllabi, exams, lesson plans) with sub-agent orchestration | Strands | AgentCore |
| Echo Prepare | Helps students study — web research, practice questions, confidence tracking | Strands | AgentCore |
| Monitoring | CloudWatch log/metric queries via MCP Gateway | Strands | AgentCore |
| Video Agent | Video metadata, transcripts, engagement analytics | Strands | AgentCore |
| Documents Agent | Document context retrieval, analytics, session management | Strands | AgentCore |
-
AWS Account with appropriate permissions — Create AWS Account
-
AWS CLI installed and configured — Install
aws configure set region us-west-2 -
Python 3.8+
-
uv package manager — Install guide
-
API Keys:
- Tavily API Key: tavily.com (for Echo Prepare web search)
-
Supported Regions:
Region Code Region Name Status us-west-2US West (Oregon) ✅ Supported
# Clone the repository
git clone https://github.com/aaravmat1209/agent_orchestration_echo.git
cd agent_orchestration_echo
# Run the interactive deployment script (works on CloudShell)
bash deploy.shThe deployment script handles: AWS CLI verification, credential checks, parameter collection, S3 bucket creation, and parallel CloudFormation stack deployment (~10-15 min).
cd frontend
npm install
chmod +x ./setup-env.sh
./setup-env.sh
npm run dev# Test individual agents
uv run test/connect_agent.py --agent monitor
uv run test/connect_agent.py --agent hostuv run monitoring_strands_agent/scripts/get_m2m_token.py
uv run echo_ink_agent/scripts/get_m2m_token.py
uv run echo_prepare_agent/scripts/get_m2m_token.py
uv run video_strands_agent/scripts/get_m2m_token.py
uv run documents_strands_agent/scripts/get_m2m_token.py
uv run host_strands_agent/scripts/get_m2m_token.pyUse the A2A Inspector to debug and validate A2A communication. Paste the agent URL and bearer token (Bearer <token>) along with headers Authorization, X-Amzn-Bedrock-AgentCore-Runtime-Session-Id (≥32 chars), and X-Amzn-Bedrock-AgentCore-Runtime-Custom-Actorid.
bash cleanup.shDeletes all stacks in reverse order, empties S3 buckets, and removes AgentCore resources (~10-15 min).
Warning
This permanently deletes all deployed resources. Cannot be undone.