Deploy 5-Agent A2A System with ADK, Gemini CLI on Lightsail
Clone repo, use pyenv (Python 3.13.13), nvm, Gemini CLI skills, and Makefile to build/test/deploy multi-agent app (Researcher/Judge/Orchestrator/Content/Course Builders) locally then to AWS Lightsail.
Streamline Multi-Agent Dev with pyenv, nvm, and Gemini CLI Skills
Manage Python versions consistently across platforms using pyenv to install Python 3.13.13—run python --version to verify—avoiding deployment issues from version mismatches common in AI/ML workflows. Pair with nvm for Node.js stability required by Gemini CLI (install via npm install -g @google/gemini-cli), authenticating via Google account for real-time code assistance. Activate ADK-specific Gemini CLI skills like adk-cheatsheet (API refs for agents/tools/orchestration), adk-scaffold (new projects/RAG additions), adk-deploy-guide (Cloud Run/GKE/CI/CD), and adk-eval-guide (metrics/LLM judging)—list via /skills list—to accelerate debugging and productionize agents faster than generic prompts. Clone https://github.com/xbill9/gemini-cli-aws, source init2.sh or set_env.sh for env vars like PROJECT_ID, then make install to handle root/agents/frontend deps.
Verify and Run Agents Locally Before Scaling
Test single agent with adk run researcher/ using Gemini 2.5-flash model, confirming logs at /tmp/agents_log/agent.latest.log and session storage in SQLite. Launch web UI via adk web --host 0.0.0.0 --allow_origins 'regex:.*' (for Cloud Shell CORS) at http://0.0.0.0:8000 to interact visually. Start full 5-agent system (Researcher on 8001, Judge 8002, Orchestrator 8004, etc., frontend 5173/backend 8000) with make start, monitoring via make local-status which checks ports/processes—stops cleanly with make stop. Run make test (pytest), make lint (ruff), or e2e-test against localhost to catch issues early, extending a Google Codelab for A2A protocol multi-agent orchestration.
Automate Lightsail Deployment for Production Workloads
Leverage AWS Lightsail VPS (pre-configured compute/storage/networking at low fixed cost) for simple hosting—access console at https://lightsail.aws.amazon.com/ls/webapp/home/containers. Use Makefile targets like deploy-lightsail to push all services, lightsail-status/endpoint-lightsail to monitor, and destroy-lightsail for cleanup, handling full lifecycle without manual scripting. This testbed—updated from Codelab with Gemini CLI—enables incremental dev: local build/debug via ADK CLI/web, then one-command cloud deploy, ideal for Python-based ADK agents treating AI like modular software engineering with state/tools like Google Search.