Lean Agentic AI on Google Cloud: Build Agents That Earn Their Watts

What does it really cost to run an AI agent at production scale?

Not only in API charges, but also in energy, carbon, water, and infrastructure.

In this presentation, Navveen Balani introduces Lean Agentic AI, a practical framework for building AI agents that deliver the required outcome using fewer computational and physical resources.

Using an ESG portfolio-rebalancing agent on Google Cloud as a worked example, the session explores how to optimize all six stages of an agentic workflow:

Goal ingestion Planning and reasoning Tool and model selection Action execution Reflection and retry Memory and learning

The presentation covers practical techniques including model routing, bounded agent graphs, structured outputs, context caching, serverless execution, batch processing, confidence-based retries, selective memory, carbon-aware scheduling, and task-level observability.

It also shows how the Software Carbon Intensity specification can help measure the carbon impact of an agent task and connect sustainability with cost, latency, and quality.

The central message is simple:

You wouldn’t send a Ferrari to deliver a pizza. Think big. Design and deploy lean.

Presented by Navveen Balani Author of Lean Agentic AI

#AgenticAI #LeanAI #GoogleCloud #Gemini #SustainableAI #GreenSoftware #AIEngineering #GenerativeAI #CloudComputing