devLap

Multi-Agent A2A with the Agent Development Kit(ADK), Amazon Fargate, and Gemini CLI

dev.to

Leveraging the Google Agent Development Kit (ADK) and the underlying Gemini LLM to build Multi-Agent Applications with A2A protocol support using the Python programming language deployed to AWS Fargate.

Aren’t There a Billion Python ADK Demos?

Yes there are.

Python has traditionally been the main coding language for ML and AI tools. The goal of this article is to provide a multi-agent test bed for building, debugging, and deploying multi-agent applications.

Rock and roll ain’t noise pollution

So what is different about this lab compared to all the others out there?

This is one of the first deep dives into a Multi-Agent application leveraging the advanced tooling of Gemini CLI. The starting point for the demo was an existing Codelab- which was updated and re-engineered with Gemini CLI.

The original Codelab- is here:

Building a Multi-Agent System | Google Codelabs

Python Version Management

One of the downsides of the wide deployment of Python has been managing the language versions across platforms and maintaining a supported version.

The pyenv tool enables deploying consistent versions of Python:

GitHub - pyenv/pyenv: Simple Python version management

As of writing — the mainstream python version is 3.13. To validate your current Python:

python --version
Python 3.13.13
Enter fullscreen mode Exit fullscreen mode

Amazon Fargate

AWS Fargate is a serverless, pay-as-you-go compute engine for containers that works with Amazon Elastic Container Service (ECS) or Elastic Kubernetes Service (EKS). It eliminates the need to manage, patch, or scale underlying EC2 virtual machines. Fargate automatically allocates, scales, and manages compute infrastructure, allowing developers to focus solely on designing and operating applications.

Details are here:

Serverless Compute - AWS Fargate - AWS

Gemini CLI

If not pre-installed you can download the Gemini CLI to interact with the source files and provide real-time assistance:

npm install -g @google/gemini-cli
Enter fullscreen mode Exit fullscreen mode

Testing the Gemini CLI Environment

Once you have all the tools and the correct Node.js version in place- you can test the startup of Gemini CLI. You will need to authenticate with a Key or your Google Account:

▝▜▄ Gemini CLI v0.33.1
    ▝▜▄
   ▗▟▀ Logged in with Google /auth
  ▝▀ Gemini Code Assist Standard /upgrade no sandbox (see /docs) /model Auto (Gemini 3) | 239.8 MB
Enter fullscreen mode Exit fullscreen mode

Node Version Management

Gemini CLI needs a consistent, up to date version of Node. The nvm command can be used to get a standard Node environment:

GitHub - nvm-sh/nvm: Node Version Manager - POSIX-compliant bash script to manage multiple active node.js versions

Agent Development Kit

The Google Agent Development Kit (ADK) is an open-source, Python-based framework designed to streamline the creation, deployment, and orchestration of sophisticated, multi-agent AI systems. It treats agent development like software engineering, offering modularity, state management, and built-in tools (like Google Search) to build autonomous agents.

The ADK can be installed from here:

Agent Development Kit (ADK)

Agent Skills

Gemini CLI can be customized to work with ADK agents. Both an Agent Development MCP server, and specific Agent skills are available.

More details are here:

Agent Development Kit (ADK)

To get the Agent Skills in Gemini CLI:

> /skills list
Available Agent Skills:
Enter fullscreen mode Exit fullscreen mode

and the ADK documentation:

> /mcp list
Configured MCP servers:
🟢 adk-docs-mcp (from adk-docs-ext) - Ready (2 tools)
  Tools:
  - mcp_adk-docs-mcp_fetch_docs
  - mcp_adk-docs-mcp_list_doc_sources
Enter fullscreen mode Exit fullscreen mode

Where do I start?

The strategy for starting multi agent development is a incremental step by step approach.

First, the basic development environment is setup with the required system variables, and a working Gemini CLI configuration.

Then, ADK Multi-Agent is built, debugged, and tested locally. Finally — the entire solution is deployed to AWS Fargate.

Setup the Basic Environment

At this point you should have a working Python environment and a working Gemini CLI installation. All of the relevant code examples and documentation is available in GitHub.

The next step is to clone the GitHub repository to your local environment:

cd ~
git clone https://github.com/xbill9/gemini-cli-aws
cd multi-fargate
Enter fullscreen mode Exit fullscreen mode

Then run init2.sh from the cloned directory.

The script will attempt to determine your shell environment and set the correct variables:

source init2.sh
Enter fullscreen mode Exit fullscreen mode

If your session times out or you need to re-authenticate- you can run the set_env.sh script to reset your environment variables:

source set_env.sh
Enter fullscreen mode Exit fullscreen mode

Variables like PROJECT_ID need to be setup for use in the various build scripts- so the set_env script can be used to reset the environment if you time-out.

Finally install the packages and dependencies:

make install
Enter fullscreen mode Exit fullscreen mode

Verify The ADK Installation

To verify the setup, run the ADK CLI locally with the researcher agent:

xbill@penguin:~/gemini-cli-aws/multi-fargate/agents$adkrunresearcher/home/xbill/.local/lib/python3.13/site-packages/google/adk/features/_feature_decorator.py:72:UserWarning:[EXPERIMENTAL]featureFeatureName.PLUGGABLE_AUTHisenabled.check_feature_enabled()Logsetupcomplete:/tmp/agents_log/agent.20260412_164250.logToaccesslatestlog:tail-F/tmp/agents_log/agent.latest.log{"asctime":"2026-04-12 16:42:50,986","name":"root","levelname":"INFO","message":"Logging initialized for researcher","filename":"logging_config.py","lineno":54,"service":"researcher","log_level":"INFO"}{"asctime":"2026-04-12 16:42:50,987","name":"researcher.agent","levelname":"INFO","message":"Initialized researcher agent with model: gemini-2.5-flash","filename":"agent.py","lineno":85}{"asctime":"2026-04-12 16:42:50,988","name":"google_adk.google.adk.cli.utils.envs","levelname":"INFO","message":"Loaded .env file for researcher at /home/xbill/gemini-cli-aws/multi-eks/.env","filename":"envs.py","lineno":83}{"asctime":"2026-04-12 16:42:50,988","name":"google_adk.google.adk.cli.utils.local_storage","levelname":"INFO","message":"Using per-agent session storage rooted at /home/xbill/gemini-cli-aws/multi-eks/agents","filename":"local_storage.py","lineno":84}{"asctime":"2026-04-12 16:42:50,988","name":"google_adk.google.adk.cli.utils.local_storage","levelname":"INFO","message":"Using file artifact service at /home/xbill/gemini-cli-aws/multi-eks/agents/researcher/.adk/artifacts","filename":"local_storage.py","lineno":110}{"asctime":"2026-04-12 16:42:50,988","name":"google_adk.google.adk.cli.utils.service_factory","levelname":"INFO","message":"Using in-memory memory service","filename":"service_factory.py","lineno":266}{"asctime":"2026-04-12 16:42:50,993","name":"google_adk.google.adk.cli.utils.local_storage","levelname":"INFO","message":"Creating local session service at /home/xbill/gemini-cli-aws/multi-eks/agents/researcher/.adk/session.db","filename":"local_storage.py","lineno":60}Runningagentresearcher,typeexittoexit.[user]:
Enter fullscreen mode Exit fullscreen mode

Test The ADK Web Interface

This tests the ADK agent interactions with a browser:

xbill@penguin:~/gemini-cli-aws/multi-fargate/agents$ adk web --host 0.0.0.0
/home/xbill/.local/lib/python3.13/site-packages/google/adk/features/_feature_decorator.py:72: UserWarning: [EXPERIMENTAL] feature FeatureName.PLUGGABLE_AUTH is enabled.
  check_feature_enabled()
2026-04-12 16:43:14,152 - INFO - service_factory.py:266 - Using in-memory memory service
2026-04-12 16:43:14,153 - INFO - local_storage.py:84 - Using per-agent session storage rooted at /home/xbill/gemini-cli-aws/multi-eks/agents
2026-04-12 16:43:14,153 - INFO - local_storage.py:110 - Using file artifact service at /home/xbill/gemini-cli-aws/multi-eks/agents/.adk/artifacts
/home/xbill/.local/lib/python3.13/site-packages/google/adk/cli/fast_api.py:198: UserWarning: [EXPERIMENTAL] InMemoryCredentialService: This feature is experimental and may change or be removed in future versions without notice. It may introduce breaking changes at any time.
  credential_service = InMemoryCredentialService()
/home/xbill/.local/lib/python3.13/site-packages/google/adk/auth/credential_service/in_memory_credential_service.py:33: UserWarning: [EXPERIMENTAL] BaseCredentialService: This feature is experimental and may change or be removed in future versions without notice. It may introduce breaking changes at any time.
  super(). __init__ ()
INFO: Started server process [32675]
INFO: Waiting for application startup.

+-----------------------------------------------------------------------------+
| ADK Web Server started |
| |
| For local testing, access at http://0.0.0.0:8000. |
+-----------------------------------------------------------------------------+

INFO: Application startup complete.
INFO: Uvicorn running on http://0.0.0.0:8000 (Press CTRL+C to quit)
Enter fullscreen mode Exit fullscreen mode

Then use the web interface — either on the local interface 127.0.0.1 or the catch-all web interface 0.0.0.0 -depending on your environment:

Special note for Google Cloud Shell Deployments- add a CORS allow_origins configuration exemption to allow the ADK agent to run:

adk web --host 0.0.0.0 --allow_origins 'regex:.*'
Enter fullscreen mode Exit fullscreen mode

Multi Agent Design

The multi-agent deployment consists of 5 agents:

  • Researcher
  • Judge
  • Orchestrator
  • Content Builder
  • Course Builder

For a detailed analysis of the multi-agent architecture- this article provides the background information:

Multi-Agent A2A with the Agent Development Kit(ADK), Cloud Run, and Gemini CLI

Running/Testing/Debugging Locally

The main Makefile has been extended with extensive targets for managing the agents on the local development environment.

First check for local running agents:

✦ I will run the make local-status command to show you the current status of all local services.

╭────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────╮
│ ✓ Shell Checking the status of local processes. │
│ │
│ --- Local Process Status --- │
│ Service Port Status PID │
│ Frontend 5173 STOPPED - │
│ Backend 8000 STOPPED - │
│ Researcher 8001 STOPPED - │
│ Judge 8002 STOPPED - │
│ Builder 8003 STOPPED - │
│ Orchestrator 8004 STOPPED - │
│ │
╰────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯

✦ All local processes are currently STOPPED.
Enter fullscreen mode Exit fullscreen mode

Then all the agents can be started together:

xbill@penguin:~/gemini-cli-aws/multi-fargate/agents$ make start
Stopping any existing agent and server processes...
Starting all agents in background...
Waiting for sub-agents to start...
All agents started. Logs: researcher.log, judge.log, content_builder.log, orchestrator.log
Starting App Backend in background...
Starting Frontend dev server in background...
All services started. Logs: researcher.log, judge.log, content_builder.log, orchestrator.log, backend.log, frontend.log
Frontend: http://localhost:5173
Backend: http://localhost:8000

make local-status again to verify.

╭────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────╮
│ ✓ Shell Verifying corrected local-status with updated pattern. │
│ │
│ --- Local Process Status --- │
│ Service Port Status PID │
│ Frontend 5173 STOPPED - │
│ Backend 8000 RUNNING 20045 │
│ Researcher 8001 RUNNING 19668 │
│ Judge 8002 RUNNING 19673 │
│ Builder 8003 RUNNING 19676 │
│ Orchestrator 8004 RUNNING 20037 │
│ │
╰────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯

✦ The backend is now showing as RUNNING, but the frontend has reverted to STOPPED. It appears the Vite process is exiting after its initial start.
  I'll check the frontend.log again.
Enter fullscreen mode Exit fullscreen mode

Then connect to the front end locally:

And the entire agent system will run in the local environment:

Local Logging / Debugging

Gemini CLI has full access to the local agent logs for debugging and troubleshooting:

> review the local logs for the multi-agent run
✦ I will check run_local.sh and the project structure to determine where logs are stored or how they are generated.
Enter fullscreen mode Exit fullscreen mode

This will verify the local operation of the code:

✦ The local logs confirm a successful multi-agent run for the topic "the history of the internet." Here is a summary of the execution flow:

* Orchestrator (orchestrator.log): Successfully managed the pipeline. It captured the topic, coordinated the Research-Judge loop, and escalated to
     content generation after the first iteration.
   * Researcher (researcher.log): Gathered approximately 9.2KB of research data using the google_search tool. It completed its task with a 200 OK
     response.
   * Judge (judge.log): Evaluated the findings and gave a "pass" status. The judge noted that the research was "comprehensive, accurate, and
     well-structured," making it an "excellent foundation for educational content."
   * Content Builder (content_builder.log): Successfully transformed the validated research into a structured Markdown course module and finished
     execution.
   * Backend & Frontend (backend.log, frontend.log): The FastAPI backend (port 8000) and Vite frontend (port 5173) are running and communicating with
     the orchestrator as expected.

Notable Observation:
   * A minor warning appeared in researcher.log: Researcher callback: 'response' not found or has no parts. However, this did not impact the
     successful completion of the research or the overall pipeline.

All services are currently active and logged at the DEBUG level as specified in run_local.sh.
Enter fullscreen mode Exit fullscreen mode

Deploying to Amazon Fargate

The project level Makefile has targets for managing the Agent deployment to serverless endpoints. A utility script checks the deployment status:

make status
Enter fullscreen mode Exit fullscreen mode

You can submit the build for GKE deployment:

✦ The application has been successfully deployed to AWS Fargate. All microservice images (Researcher, Judge, Content Builder, Orchestrator, and
  App) have been built and pushed to Amazon ECR, and the ECS service has been updated.
Enter fullscreen mode Exit fullscreen mode

Once the containers are deployed- you can then get the endpoint:

make endpoint, to retrieve the public IP of your Fargate service.

╭────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────╮
│ ✓ Shell Retrieving the public endpoint for the Fargate service. │
│ │
│ Fargate Endpoint: http://54.221.168.167:8080 │
│ │
╰────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯

✦ The public endpoint for your Fargate service is: http://54.221.168.167:8080

  You can use this URL to access your application or run an end-to-end test against it with make e2e-test-fargate.
Enter fullscreen mode Exit fullscreen mode

The status can then be checked:

> make status
✦ I will check the status of the AWS Fargate services.

╭────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────╮
│ ✓ Shell Checking the status of the AWS Fargate services. │
│ │
│ ----------------------------------------------------------------------------------------------------------------------------------------- │
│ | DescribeServices | │
│ +---------+----------------------+----------+---------+---------------------------------------------------------------------------------+ │
│ | Desired | Name | Running | Status | TaskDef | │
│ +---------+----------------------+----------+---------+---------------------------------------------------------------------------------+ │
│ | 1 | adk-fargate-service | 1 | ACTIVE | arn:aws:ecs:us-east-1:106059658660:task-definition/adk-course-creator-task:2 | │
│ +---------+----------------------+----------+---------+---------------------------------------------------------------------------------+ │
│ │
╰────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯

✦ The AWS Fargate service adk-fargate-service is currently ACTIVE with 1 running task.
Enter fullscreen mode Exit fullscreen mode

The service will be visible in the AWS console:

And the entire system can be tested:


✦ The end-to-end test against the deployed AWS Fargate endpoint was successful! The system correctly orchestrated the microservices to create a
  structured "History of the Internet" course and streamed the results back via Server-Sent Events (SSE).

  Deployment & Verification Summary:
   - Status: The adk-fargate-service is ACTIVE and running.
   - Endpoint: http://54.221.168.167:8080
   - E2E Test: Completed successfully against the remote Fargate task.
   - Local Monitoring: make local-status is available and correctly reports local process states using robust pattern matching.

  Your multi-agent AI Course Creator is now fully deployed and operational on AWS Fargate.
Enter fullscreen mode Exit fullscreen mode

Running the Web Interface

Start a connection to the Cloud Run deployed app:

Then connect to the app :

Then use online course generator:

Review Fargate Setup

Gemini CLI was used to review the Fargate Setup:

  ✓ ReadFile setup-fargate.sh

✦ All agents in this project are configured to run within a single AWS Fargate task. Therefore, they share a common pool of resources:

   * Total Task CPU: 1024 (1 vCPU)
   * Total Task Memory: 2048 (2 GB RAM)

  Components Sharing these Resources:
  The following five containers share the 1 vCPU and 2 GB RAM:
   1. app: The FastAPI backend and web interface.
   2. orchestrator: Manages the agent workflow.
   3. researcher: Performs topic research using Google Search.
   4. judge: Evaluates research findings.
   5. content-builder: Generates the final course modules.

  This configuration is defined in the setup-fargate.sh script, which registers the ECS task definition with these specific resource allocations.
Enter fullscreen mode Exit fullscreen mode

and a deep dive into the deployment pattern:

> how is this app deployed on Fargate?
✦ I will examine the deploy-fargate.sh script and the root Makefile to understand the complete build and push process for the Fargate deployment.

  ✓ ReadFile deploy-fargate.sh

✦ The deployment of the AI Course Creator to AWS Fargate is a multi-step process managed by two primary scripts: setup-fargate.sh and
  deploy-fargate.sh.

  1. Infrastructure Provisioning (setup-fargate.sh)
  This script sets up the foundational AWS resources required to run the application:
   * ECS Cluster: Creates a cluster named adk-fargate-cluster.
   * IAM Roles: Provisions the ecsTaskExecutionRole-adk with the necessary permissions to pull images from ECR and send logs to CloudWatch.
   * Security Groups: Configures an adk-fargate-sg that allows inbound traffic on port 8080 (the web app port).
   * Task Definition: Defines how the 5 microservices (app, orchestrator, researcher, judge, content-builder) are run. It specifies a single
     Fargate task where all 5 containers share the same network stack (awsvpc) and resource pool (1 vCPU, 2 GB RAM). This allows them to
     communicate with each other using localhost.
   * ECS Service: Creates a Fargate service that maintains one running instance of the task definition.

  2. Build and Push Process (deploy-fargate.sh)
  This script handles the continuous delivery of the application code:
   * ECR Repositories: Ensures that Amazon Elastic Container Registry (ECR) repositories exist for each of the 5 microservices.
   * Authentication: Logs in the local Docker daemon to the ECR registry using the AWS CLI.
   * Container Builds: Iteratively builds Docker images for each service:
       * researcher
       * judge
       * content-builder
       * orchestrator
       * app (FastAPI backend + Frontend)
   * Image Push: Tags the built images and pushes them to their respective ECR repositories.
   * Deployment Trigger: Once the images are pushed, the ECS service can be updated (e.g., via aws ecs update-service --force-new-deployment) to
     pull the latest versions and restart the tasks.

  Summary of Deployment Flow
   1. Run setup-fargate.sh (once) to create the AWS infrastructure.
   2. Run deploy-fargate.sh to build and upload the container images.
   3. Monitor Status: Use make status-fargate (which likely wraps aws ecs describe-tasks) to verify the containers have successfully reached the
      RUNNING state.

  This architecture leverages Fargate's serverless nature to run a complex, multi-agent system without managing the underlying EC2 instances,
  while the single-task approach simplifies service discovery and reduces cross-service latency.
Enter fullscreen mode Exit fullscreen mode

Final Gemini CLI Code Review

As a final step — Gemini CLI was used for a full code review of the project:

✦ I've analyzed the project's architecture, security, and agent logic. I'll now provide a code review of the AI Course Creator project, focusing
  on its distributed multi-agent design using ADK and A2A on AWS Fargate.

  1. Architecture & Orchestration
   * A2A Workflow: The system uses a clean SequentialAgent pipeline (course_creation_pipeline) in the orchestrator to chain specialized agents.
   * State Management: The StateCapturer pattern is an effective way to bridge the context between independent A2A services. It intercepts LLM
     responses and injects them into the session_state, which subsequent agents then consume.
   * Loop Logic: Using LoopAgent with a max_iterations=2 limit for the Research-Judge cycle is a best practice to prevent infinite loops and
     control costs.
   * Escalation Control: The EscalationChecker correctly uses structured Pydantic output (JudgeFeedback) to decide when to break the research
     loop.

  2. ADK Best Practices
   * Model Selection: The project correctly prioritizes gemini-2.5-flash for efficiency and reasoning, with gemini-2.5-pro as an alternative for
     complex tasks.
   * Tool Usage: The google_search tool in the researcher agent is well-integrated.
   * Structured Outputs: Using Pydantic schemas for the judge agent ensures that the orchestration logic is robust and type-safe.

  3. Security & Authentication
   * Service-to-Service Auth: The authenticated_httpx.py utility is a standout feature. It handles Google OIDC tokens for secure communication
     between agents, even when deployed on non-Google infrastructure like AWS Fargate.
   * Environment Variables: Sensitive information like GOOGLE_API_KEY is passed via environment variables, following standard security practices.
   * Dynamic A2A Cards: The a2a_utils.py middleware solves the challenge of discovering service URLs in dynamic container environments by
     rewriting the Agent Card's url based on the x-forwarded-host header.

  4. Infrastructure & Deployment
   * Resource Efficiency: Running all 5 containers in a single Fargate task (1 vCPU, 2 GB RAM) is cost-effective for a prototype but may lead to
     resource contention if multiple courses are generated simultaneously.
   * Docker Strategy: Using a standardized shared/adk_app.py entry point across all agents ensures consistent logging, A2A registration, and
     middleware application.

  Summary
  The project demonstrates a high level of technical maturity in distributed agent orchestration. It effectively adapts Google ADK patterns for a
  cross-cloud environment (AWS Fargate) while maintaining strong security and clean architectural boundaries.
Enter fullscreen mode Exit fullscreen mode

Summary

The Agent Development Kit (ADK) was used to build a multi-agent system with A2A support using the Gemini Flash LLM Model. This application was tested locally with Gemini CLI and then deployed to AWS Fargate. Finally, Gemini CLI was used for a complete project code review.

Source: dev.to

arrow_back Back to News