Lean Agentic AI at Google I/O

It was a pleasure speaking at the Google I/O event on Lean Agentic AI.

The agentic gold rush — opening the Lean Agentic AI session at Google I/O.

As organizations move from deploying a handful of AI agents to thousands, efficiency can no longer be an afterthought. We need to design agentic systems that are efficient by design — across cost, energy, and carbon.

The agentic gold rush

I opened with the numbers that frame the moment: over 6 trillion tokens a month already flow through Gemini via the Agent Development Kit; data-centre electricity demand is projected to grow +128% — from roughly 415 TWh in 2024 toward 945 TWh by 2030; and with Agent Runtime, agents can now run for 7 days without sleeping. Enterprises are deploying agents at unprecedented speed — but the bill, in money and in energy, is only just starting to land.

What the session covered

Efficiency as an architecture concern, not an afterthought. Every design decision in an agentic system — which model handles which step, how much context travels with each call, how many times an agent retries, how long it runs — is simultaneously a cost, energy, and carbon decision. Lean Agentic AI treats those three as first-class design inputs, the same way we treat latency and reliability.

Lean principles for agent design. The core of the session walked through practical patterns: right-sizing models to the task instead of defaulting to the largest; routing requests by difficulty; giving every agent an explicit budget for tokens, loops, and retries; caching what agents repeatedly recompute; and knowing when a long-running agent should stop. Small decisions, multiplied by thousands of agents, decide whether the system scales or the bill does.

Measure what matters, then engineer for reduction. You cannot optimize what you do not measure — which is where standardized measurement comes in.

The conversation after the session

One of the most rewarding parts was the discussion with developers after the session. Many hadn’t considered that every architectural decision in an AI system has implications beyond performance — it also affects cost, energy, and carbon. Raising that awareness is important because sustainability isn’t just about reducing environmental impact; it’s also one of the most effective ways to lower the total operational cost of AI. That’s how we make sustainable AI the new engineering norm.

SCI for AI

As part of the Green Software Foundation (GSF), I’m also excited about the progress we’re making with the SCI for AI standard. Rather than focusing on offsets, SCI for AI provides a standardized way to measure, compare, and continuously reduce the carbon impact of AI workloads. The emphasis is simple: measure what matters, then engineer for reduction.

Slides and demo

Many attendees requested the slides and demo after the session, so I’ve made them available on GitHub: Lean Agentic AI — slides and demo

I’d love to hear your thoughts and continue the conversation on building lean, efficient, and sustainable agentic AI.