Posts

Tomorrow, I’ll be speaking at Google I/O Connect, Bengaluru on a topic I’ve become deeply passionate about: Lean Agentic AI.

We’re entering an era where organizations will deploy not one or two agents—but thousands. The question is no longer: can we build them?

It’s: can we design agents that are efficient by design — across cost, energy, and carbon?

In this session, I’ll show how to engineer production-ready agentic systems that optimize across three currencies that compound simultaneously:
💰 Dollars — reducing operational cost
⚡ Watts — improving energy efficiency
🌍 Carbon — attaching a real gCO₂e number to every agent task using the Green Software Foundation’s Software Carbon Intensity (SCI) ISO specification.

Every wasted token costs money, watts, and emissions at once.

But carbon is only part of the story. As AI adoption accelerates, so do its impacts on water consumption, e-waste, and grid demand.

We’ll also explore why measurement is the missing layer—connecting cost, energy, carbon, water, and waste into a single engineering discipline so we can optimize what actually matters.

We’ll walk through practical architectural patterns on the Google Gemini Enterprise Agent Platform: model selection, orchestration, observability, evaluation, simulation, memory, and carbon-aware execution.

The goal is not just measurement—it’s continuous reduction of cost, energy, carbon, water, and waste.

If you’re attending Google I/O Connect, I’d love for you to join me at Community Lounge 1. And if you’re around afterwards, please come and say hello. I’d love to chat about Agentic AI, Green Software, or whatever you’re building.

More than anything, I’d love to see our developer community start treating cost and environmental impact as first-class engineering metrics.

A big thank you to Paul Ravindranath G , Rajat Bhatia, and the Google I/O Connect team for the invitation.

See you there!

#GoogleIO #GoogleIOConnect #BuildWithGemini #GoogleCloud #GeminiEnterprise #Gemini #AgenticAI #LeanAgenticAI #GreenSoftware #AIEngineering #Sustainability #FinOps #SoftwareEngineering Green Software Foundation

🚨 AI is advancing faster than your organization can transform.

That may be the single biggest reason why so many enterprises are struggling to demonstrate ROI from Agentic AI.

Every month, we see smarter models, better reasoning, lower costs, and more capable agents. Yet many business leaders are still asking the same question:

We’ve invested in AI… so where’s the business value?

The problem isn’t that AI isn’t ready.
The problem is that AI evolves at machine speed, while organizations transform at human speed.

In the latest edition of Technology Bytes, I explore what I call The Transformation Gap—the growing distance between rapidly evolving AI capabilities and the slower pace of organizational change.

The article also explores why:

• Agentic AI introduces a new economic model built on tokenomics, where development, testing, and innovation all consume compute.
• Enterprises are no longer managing static software—they’re managing continuously evolving intelligence.
• Benchmark scores don’t create business value. Redesigned business processes do.
• The organizations that achieve the greatest ROI won’t necessarily deploy the smartest agents—they’ll redesign how work gets done.

The future of enterprise AI isn’t just about building better models.
It’s about building organizations that can evolve alongside them.

📖 Read the latest Technology Bytes: Agentic AI ROI — Why Most Organizations Aren’t Seeing Business Value Yet

I’d love to hear your perspective:

What’s the biggest barrier to realizing ROI from Agentic AI today—technology, governance, organizational change, or something else?

Day 19 of Green, Efficient AI is live — The Test Tail.

A team spent six months bringing production AI costs down. The playbook worked — production inference became dramatically cheaper. The monthly bill barely moved.

The gap was hiding in plain sight. Every merge triggered a full eval sweep. The judge model was pricier than the model being judged. Prompt sweeps ran untracked, most discarded within the hour.

Meter. Gate. Sample. Substitute. The ladder that turns development inference from an unbudgeted reflex into a governed line on the bill.

Read the full issue ↓

Green Software Foundation #greenai #efficentai #aieconomics #leanagenticai

Day 18 of Green, Efficient AI is live — The Forever Data.

Every AI system carries two footprints. The work it performs, and the data it leaves behind. The second one rarely gets named.

A B2B platform running a retrieval assistant found that three routine embedding-model upgrades — each a drop-in change on the API side — had quietly become one of the largest non-production AI infrastructure expenses on the account that year. Nobody had put re-embedding on the upgrade review.

Inventory. Tag. Retire. Reuse. The ladder that turns AI-scale storage from an accident into a decision.

Read the full issue ↓

Green Software Foundation #greenai #efficentai #leanagentciai #aieconomics

AI doesn’t need another breakthrough model.
It needs a return to engineering fundamentals.

Somewhere between foundation models and agentic AI, we started confusing capability with architecture. Bigger models, more agents, longer context windows, and increasingly complex workflows aren’t always better engineering.

In this edition of Technology Bytes, I explore why the next competitive advantage in AI may not come from building more intelligent systems—but from building better-engineered ones.

And why it’s time to move from the “reel economy” of impressive AI demos to the real economy of measurable business outcomes.

📖 Read the full article below.

Day 17 of Green, Efficient AI is live — The Invisible Payload.

A user asks a support agent to reset their password — seven words, roughly ten tokens. The trace shows the model actually receives thousands of tokens. A system prompt written months ago. Every tool schema in the catalog, whether the tool is needed or not. Few-shot examples fixing behaviours long since stabilised. Formatting rules. Safety framing. The user’s question is at the bottom.

The prompt the model reads is not the prompt the user wrote. And the difference between the two is what shows up on the bill.

Four rungs inside: Inventory, Trim, Split, Amortise.

Read the full issue in today’s edition

Green Software Foundation #greenai #efficentai #aieconomics #leanagenticai

Day 16 of Green, Efficient AI is live — Every Guardrail Is a Call.

The most overlooked line item in an AI system is the safety wrap. So why do most teams run four or five model-based checks on every request without ever costing them?

In many production pipelines, the answer is design-time muscle memory, not architecture. A customer-facing assistant I describe in this issue was running four model-based safety checks per call — and the wrap was consistently consuming more tokens than the assistant itself.

Day 16 lays out a four-rung ladder — Scope, Tier, Cache, Escalate — for paying for risk, not for routine.

Read the full issue.

Green Software Foundation #greenai #leanagenticai #aieconomics

The future of enterprise AI is not about maximizing intelligence. It is about delivering the minimum intelligence required to achieve the maximum business outcome.

This idea has increasingly shaped how I think about the next phase of enterprise AI.

Over the past few years, we’ve witnessed remarkable advances in frontier AI models. The pace of innovation has been extraordinary, fundamentally changing what organizations believe is possible.

But as enterprises move from experimentation to production, a new set of questions is emerging:
➡️ How do we operationalize intelligence at scale?
➡️ How do we govern autonomous agents?
➡️ How do we ensure security, compliance, and data sovereignty?
➡️ How do we optimize for cost, latency, energy, and carbon?
➡️ How do we translate frontier intelligence into measurable business outcomes?

In the latest edition of the Technology Bytes newsletter, I explore why the next competitive frontier in enterprise AI is not simply more intelligence.

It’s about integrating intelligence into enterprise systems, governing it responsibly, and optimizing it for outcomes.

I also introduce the concept of Intelligence Economics — measuring AI not just by tokens consumed, but by outcomes delivered per dollar, per joule, per gram of CO₂, and per second of latency.

As frontier intelligence becomes widely accessible, organizations that succeed will be those that can operationalize it most effectively and efficiently.
This is the essence of what I call Lean Agentic AI.

📖 Read the latest edition for details

What do you think will differentiate enterprise AI leaders over the next 3 years: access to frontier models, or the ability to operationalize intelligence at scale?

Day 15 of Green, Efficient AI is live — The Route Before the Model.

The model is usually the most expensive component in an AI system. So why does every request still reach it?

In many production pipelines, the answer is reflex, not architecture. A telecom team I describe in this issue discovered that the bulk of their LLM traffic was answering questions a database could have handled in milliseconds.

Day 15 lays out a four-rung ladder — Filter, Classify, Route, Call — for the route every request should travel before it ever sees the model.

Green Software Foundation #greenai #efficentai

Day 14 of Green, Efficient AI is live — The Grid You Didn’t Choose.

Most production AI runs in a region somebody picked years ago, on hardware procured for a different workload, and on schedules chosen for operational convenience rather than carbon performance.

None of those choices were carbon-aware. Most of them are still in place.

The same inference call can carry very different carbon footprints depending on where and when it runs. The model does not know. The bill is roughly the same. The atmosphere is not.

Today’s issue is about the part of the stack almost nobody owns — and one of the most underused efficiency levers in AI systems today.

Green Software Foundation #greenai #efficentai #aieconomics #leanagenticai

Last week, I had the opportunity to deliver the keynote at the CleanEnviro Summit Singapore Catalyst 2026 in Singapore, hosted by the National Environment Agency.

Dr. Janil Puthucheary, Senior Minister of State, Ministry of Sustainability and the Environment & Ministry of Education, opened the event by setting the context for the day’s discussions on sustainability. He highlighted Singapore’s progress over the past decade, including reductions in waste generation, increased household recycling participation, and greater public awareness of recycling. At the same time, he underscored the challenges ahead—from rising collection and freight costs to volatile recyclables markets and the need for innovation to extend the lifespan of Semakau Landfill beyond 2035.

It was inspiring to see Singapore’s long-term commitment to sustainability come through so clearly.

My keynote, “Innovating Towards a Sustainable Future,” built on that theme and focused on three ideas:

— Sustainability and business value are not competing priorities. Increasingly, they are the same path.
— The Environmental Services industry is moving from activity-based metrics (bins emptied, routes completed) toward outcome-based metrics (waste diverted, emissions reduced, and materials kept in circular systems). That transition represents one of the most important innovation opportunities of the next decade.
— Green AI has two dimensions. Green BY technology—using AI to make Environmental Services more sustainable. And Green IN technology—ensuring the AI itself is efficient, responsible, and sustainable. Environmental Services has a unique opportunity to advance both simultaneously from the outset.

What stayed with me most was the ecosystem represented in the room: service providers, premises owners, innovators, government agencies, investors, and academics working together on a shared challenge. The Innovation Showcase brought this collaboration to life through demonstrations spanning Autonomous Systems, Analytics, AI, and Advanced Biology—showing how quickly sustainability innovation is moving from concept to deployment.

Thank you to the National Environment Agency for the invitation, to Dr Janil Puthucheary for setting the tone, and to everyone who shared ideas and perspectives throughout the event.

As Executive Director of the Green Software Foundation, I am particularly encouraged by Singapore’s leadership in sustainable technology. Singapore was the first government to join the Green Software Foundation, demonstrating how public-sector leadership can help accelerate the adoption of sustainable technology practices and standards..

Singapore continues to demonstrate how policy, industry, and innovation can come together to accelerate sustainable outcomes.

Day 13 of Green, Efficient AI is live — The Queue You Never Built.

Most inference traffic is not waiting on a human. It is waiting on a loop.
Embeddings for an index that refreshes hourly. Classifications for a moderation queue. Summaries for a morning digest. Evaluations that run overnight.
None of these need a sub-second answer. Almost all of them are dispatched as if they did.

The GPU does not care whether it processes one request or thirty in the same forward pass. The bill does. The grid does.

Today’s issue is about the latency budget nobody set, and what changes when the question is finally asked.

Green Software Foundation #greenai #efficentai

Day 12 of Green, Efficient AI is live — Cache What Doesn’t Change.

Depending on the workload, forty to ninety-five percent of every AI prompt is the same as the last one — the system prompt, the tools, the document the agent is reading. The model reprocesses all of it from scratch, every time, paying the energy bill, the cost bill, and the latency bill twice for an output it already computed once.

The fix is not clever prompting. It is treating the prompt as a data structure with a stable part and a variable part, and telling the system which is which.

The four-rung ladder for fixing it, in today’s issue.

Green Software Foundation #greenai #efficentai #aieconomics #leanagenticai

If you’re wondering, like I was, why Anthropic’s latest model Fable suddenly shows as unavailable… this may be why.

Halfway through working on Fable, access suddenly disappeared.

The initial assumption?
A temporary issue. Capacity constraints. Service overload.
With the pace at which frontier AI models are being adopted, that felt like the obvious explanation.

But after digging deeper and reviewing the latest developments, it became clear this is part of a much broader shift.

Recent reports indicate that access to some of Anthropic’s latest frontier models, including Fable 5 and Mythos 5, has been impacted by new U.S. export-control measures tied to national security considerations.

Stepping back, this feels like an important moment for the AI ecosystem.

As AI capabilities advance, frontier models are increasingly being viewed not only as technology products, but also as strategic capabilities. It is understandable that governments will evaluate access through the lens of security, responsible deployment, and long-term societal impact.

For enterprises, countries, and builders, this reinforces an important lesson:
The future of AI strategy cannot rely on a single model or provider.

Diversity in model choices, open ecosystems, local capability building, and resilient AI architectures may become increasingly important in a world where access can evolve due to policy, regulation, or geopolitical considerations.

One more dimension for AI leaders to design for: access resilience.
Not just:
✔️ Model performance
✔️ Cost
✔️ Safety & Responsible AI
✔️ Latency & Reliability
✔️ Sustainability (Cost, Carbon, Water, Energy)

But also:
✔️ What happens if access to a frontier model changes overnight?

The strongest AI strategies may increasingly be the ones designed for optionality, adaptability, resilience, and responsible scale.

Day 11 of Green, Efficient AI is live — Reset the Loop When Context Decays.

An agent runs for fifteen turns. Each turn costs more than the one before it. By the last few turns, a single turn can cost more than the first several combined.
The questions did not get harder. The loop got heavier.

Most of what a long-running agent carries forward is residue — retracted reasoning, verbose tool payloads, instructions the agent has already satisfied.
None of it is being used on the next turn. All of it is being re-processed on every turn.

There is a crossover point where a fresh loop, started with a four-line summary, is cheaper than another turn of the old one. Knowing where that point is, in today’s issue.

Green Software Foundation #greenai #efficentai #leanagenticai

Day 10 of Green, Efficient AI is live — Close the Loop with the Right Tool.

Two agents, same customer question: “What’s the status of my last order, and when will it arrive?

Agent A has four clean tools — look up the customer, list orders, fetch shipping, fetch carrier ETA. Five turns to assemble the answer.

Agent B has one tool that already understands the question: latest order status for this customer. Two turns. Done.

Same model. Same data. The difference was whether the tools were shaped to the agent’s task or to your system’s capabilities.

Every tool an agent calls either ends a turn or feeds the next one. The interface, not the model, decides which.

The shape that closes a turn, in today’s issue.

Green Software Foundation #greenai #efficentai #leanagenticai

🚨 The Token Shock — Why Companies Are Running Out of AI Budget
🎙️ Episode 91 of Agentic AI: The Future of Intelligent Systems is out now.

Are we creating value… or simply generating activity?

That may quietly become one of the biggest questions organizations ask in the age of Agentic AI.

Because intelligence is no longer free to scale endlessly.

As copilots, coding assistants, and autonomous agents become part of everyday work, a new enterprise challenge is emerging:

Token Shock.

The moment organizations realize that AI does not behave like traditional software.

In this episode, I explore:
✅ Why organizations are beginning to experience unexpected AI cost escalation
✅ The hidden economics of tokens and reasoning
✅ Why “reasoning budgets” may become essential
✅ How invisible retries, tool usage, and agentic workflows quietly drive spend
✅ Why the future may be about smarter intelligence, not just more intelligence

Because perhaps the future winners in AI will not be the organizations consuming the most intelligence…
But the organizations using intelligence wisely.

🎧 Listen here: https://lnkd.in/dQuaivcs

More on Lean Agentic AI: leanagenticai.com

Wait… the AI budget is already gone?

This is becoming a real conversation inside organizations.

Coding assistants. Meeting summaries. Research agents. Copilots. Multi-agent workflows.

AI feels effortless… until someone sees the token bill.

In early 2025, I introduced the concept of Lean Principles for Agentic AI in my book Lean Agentic AI — focused on how organizations can apply intelligence in a lean, efficient, and purposeful way, balancing cost and environmental impact.

If you are building AI copilots, deploying agents, scaling enterprise AI, or navigating token costs, governance, and efficiency challenges, these principles and guidelines may be highly relevant.

📘 Book available worldwide on Amazon: https://amzn.to/3RQmXti
🌐 Learn more: https://leanagenticai.com/

Day 9 of Green, Efficient AI is live — Plan the Loop Before You Run It.

Two agents, same task: reconcile a billing dispute across three systems. Same model. Same tools. Same access.

One finishes in five turns. The other takes fourteen — and lands on the same answer nine turns later.

Nothing was broken. The slow one just started moving before it knew where it was going.

Without a plan, an agent decides each move from the last result it saw — never from the route as a whole. One cheap planning turn buys it sight of every step.
And it removes the turns at the start, where every wasted turn costs the most.
How to plan it, in today’s issue.

Green Software Foundation #greenai #efficentai

When did innovation become something we buy instead of something we think our way through?

The latest edition of Technology Bytes explores a question that feels increasingly relevant in the age of AI:

Has innovation shifted from ingenuity to compute?

For decades, innovation was shaped by constraints. Teams solved hard problems through creativity, efficiency, and thoughtful engineering. Today, especially in software and AI, innovation is increasingly measured through more compute, more tokens, more infrastructure, and more spend.

In this issue, I reflect on:

• Why earlier innovation often achieved more with less
• How abundance may be changing the way we solve problems
• The hidden environmental and economic costs of compute-first thinking
• Why efficiency and thoughtful design still matter in the AI era
• What it means to return innovation to its roots

A reflection on whether scale is enhancing innovation, or quietly becoming a substitute for it.