About
I have spent 25 years leading engineering organizations through the moments that decide…
Services
Articles by Anil
-
Key Leadership Roles for Building a Global Engineering Team
Key Leadership Roles for Building a Global Engineering Team
In my earlier article Building a Global Engineering Team, I wrote about elements to consider while expanding…
4 Comments -
Building a Global Engineering Team
Building a Global Engineering Team
I have been fortunate in my career to have the opportunity to build out and run successful global teams in several…
9 Comments -
'Leadership and Self Deception' & 'An Anatomy Of Peace' - an interesting read!
'Leadership and Self Deception' & 'An Anatomy Of Peace' - an interesting read!
A leadership podcast recommended this book: 'Leadership and Self Deception by The Arbinger Institute'. It is an…
4 Comments
Activity
8K followers
-
Anil Virmani shared thisMy Best Engineers Hand-Write Less Than 1% of Their Code. On Purpose. Software development just split in two, and most people have not noticed yet. Nearly every team has coding agents now. They write code, run tests, and fix failures for hours with no human watching. So you would expect everyone to be shipping faster. Most teams are not. That is the strange part, and it is the whole story. The split is not between teams that have agents and teams that do not. Almost everyone has them. The split is between people who swapped their tools and people who changed how they work. Only the second group is seeing the real gains. Here is the shift that surprised me most. The strongest engineers have mostly stopped writing software directly. Instead they build the system that writes the software. They set up the agents, the context those agents get fed, the checks that catch bad output, the whole pipeline. Once that is running well, they hand-write less than 1% of the code they ship. Sit with that number. Ninety-nine percent of the code is produced by a system they designed, not typed by them. That does not make them less skilled. It makes them skilled at a different thing. The code was never the hard part. Getting the architecture and design right was, and specifying it tightly enough that the system builds the right thing. That is still the top challenge, and agents made it harder to fake, not easier. A vague spec used to produce one bad function. Now it produces wrong code at scale, fast. The engineers who can hold the whole design in their head and pin it down precisely are the ones the machine actually multiplies. The teams stuck in the first group are working hard and wondering why the numbers have not moved. They gave their engineers a faster way to type. Then they kept every other part of the process exactly the same. The agent got faster. The workflow around it did not. So the gain leaks out everywhere the old process still runs. If you lead engineers right now, this is the real question. Are your best people still heads-down writing code, or are they building the systems that write it. The first is a good engineer in 2023. The second is where the frontier is now. The gap between them is going to decide which teams pull ahead this year.
-
Anil Virmani shared thisYour Board Thinks AI Means a Smaller Team. The Math Says Otherwise. Every founder and board I talk to is asking the same quiet question. If agents write most of the code, can I run this with half the team? I understand the instinct. But the math does not work the way it looks, and betting your org on it is a mistake. Start with what actually shrinks. Agents are good at the high-volume, well-defined coding work. So that part of the load goes down, and you need fewer pure hands-on coders for it. That is real. But that is not the whole org, and it is not even the expensive part. Here is what grows instead. Someone has to specify what to build. Someone has to check that what the agent produced is right. Someone has to design the system, own the architecture, catch the subtle failure a test would miss. That work goes up, and it is more senior, more expensive people doing it. So you do not get a smaller org. You get a different one. The center of gravity moves toward judgment and away from typing. Now the part almost everyone misses. When building gets cheaper, teams do not build the same amount for less. They build more. This is an old idea in economics called the Jevons paradox. When steam engines got more efficient, England did not burn less coal. It burned far more, because cheap power made a hundred new uses worth it. Software is the same. There is always more that people want built than any team can get to. Make building cheaper and that backlog does not sit there. It gets built. So the teams getting faster with agents are mostly not banking the savings as layoffs. They are taking on scope they would never have touched before. Same people, far more shipped. That is the winning move, and it looks nothing like a headcount cut. The trap is treating this purely as a cost lever. The orgs that do that cut the wrong roles. They thin out juniors and quality, the pipeline and the guardrails, and end up faster in the short run and fragile in the long run. Kill your junior pipeline and you have no seniors in five years. Cut your quality bar and the agents ship problems faster than ever. The right question was never how many people can I remove. It is what can this team build now that it never could before.
-
Anil Virmani shared thisYour Best People Won't Tell You AI Is Breaking the Culture. They'll Just Leave. We all know technical debt. You take a shortcut in the code to ship faster. It works. Then a year later you are paying for it in bugs and slow releases. There is a version of this for culture, and AI adds to it faster than any shift I have seen. Here is how it happens. A team starts using agents. Output goes up. Everyone is happy. But nobody stopped to redesign how the team actually works. The old norms, how decisions get made, who owns what, how people learn, how they raise a concern, all of it was built for a world where humans did the work. That world just changed. If you do not redesign those norms, they quietly break. That is culture debt, and you do not see the bill until your best people start leaving. I watched a version of this years ago, before AI, when I inherited a team in India with brutal attrition. The tooling was fine. The pay was fine. What was broken was ownership. People did the tasks handed to them and felt like cogs. They had no real say and no room to raise a problem. We fixed it by giving teams real ownership, real decisions, and real safety to speak up. Attrition dropped. The work got better. None of that was a tools problem. AI makes this risk sharper, not smaller. When a machine can do more of the work, the thing that keeps good people is not the work itself. It is whether they feel essential. Whether their judgment still matters. Whether it is safe to say "I do not trust what this agent produced" without looking like they are resisting progress. A recent IBM study found that 43% of executives said their people do not feel safe raising concerns about AI outputs. Sit with that number. Almost half of orgs have quietly told their people to keep their doubts to themselves, right when those doubts matter most. The fix is not a new tool or a training program. It is the same thing it always was. Give people real ownership. Make it safe to question the machine. Keep them learning so they feel ahead of the change, not run over by it. The companies that get this will move fast and keep their people. The ones that do not will look productive for a while, then wonder why the good ones walked out the door. Culture was never the soft part of the job. Right now it is the hardest part, and the most important.
-
Anil Virmani shared thisAgentic AI Won't Shrink Your Org. It Changes the Manager, Director, and VP Job. This is the last piece in the series. I covered the engineers, then QA and SRE. Now the management layer, where the changes are quieter and the stakes higher. Start with the manager, because the shift is subtle. You used to track what people worked on. Now you track whether the agent-driven work holds up, and whether the team is getting better at directing agents or just busier. The other new job is keeping the team learning. The tools change every few months, and the efficiency only shows up if people keep picking up the new ones. Most people ask the wrong question first, so let me answer it. Does the org shrink? Not the way you think. The mix changes more than the number. The high-volume coding work shrinks. The judgment and verification work grows. So the org leans more senior by center of gravity. That does not mean cutting juniors. Kill the junior pipeline and you have no seniors in five years. And here is the part that surprises people. When building gets cheap, teams usually build more, not less. They take on scope they would never have attempted before. The savings get spent on ambition, not cuts. So a director watching headcount as the scoreboard is watching the wrong number. That is the first real shift. A director used to fight for headcount and hand it out. Now the job is closer to capital allocation. Where do I point an org that is suddenly more powerful. Which bets do I fund. Which do I kill. How much goes to compute and tooling versus people. The second shift is owning trust across teams. One team can set its own agent standards. Across ten teams, someone has to own whether the whole org can trust what its agents produce. Shared evaluation. Common guardrails. One answer to how we know this is right. That is a director-level job now, and most orgs have not handed it to anyone.The third is the hardest. Agents make it easy for a team to look busy. Lots of commits. Lots of shipped features. Meanwhile the business outcome stalls, or production quietly rots. Your old dashboards will lie to you here, because they measure activity, and activity is what agents inflate. The hardest new job is telling real progress from fast motion. The VP owns the whole thing. Business outcomes, strategy, execution, and the health and retention of the org all sit with them. Agents raise the stakes on every one. A more powerful org pointed at the wrong problem does damage faster, so strategy matters more. Fake velocity hides real trouble, so execution matters more. And the human side matters more, not less. When the tools shift this fast, people burn out or check out, and holding a strong team together through constant change becomes core to the job. There is one thread through this whole series. Value moved from producing work to directing and verifying it. The junior engineer verifies a function. The VP verifies a business. Same job, different altitude.
-
Anil Virmani shared thisQA and SRE Are the Last Jobs Agents Will Touch. Here's Why. Last week I wrote about how agents change the engineering roles everyone talks about. Junior engineer, senior engineer, PM. The reaction that surprised me most was how many people asked about the roles nobody talks about. QA and SRE. So here they are. Start with QA, because this is where the fear is loudest. If agents write the tests and run them, what's left for a QA engineer to do? A lot, but not the part you'd expect. The old QA job was writing and running test cases against what got built. Agents do that cheaply now. What they can't do is design the system and integration tests that prove the whole thing works together, or decide what "correct" even means in the first place. Someone has to design the evaluation: the adversarial cases the agent won't think to try, the edge conditions that actually matter, the bar for what "good enough to ship" is. That work doesn't shrink. It gets harder and more valuable. QA stops being the people who run the tests and becomes the people who define how you trust a system a machine mostly built. SRE moves the same direction, for a scarier reason. Agents let you ship far more code, far faster, with fewer humans who deeply understand any single piece of it. Think about what that does to production. More change, less human familiarity with each change. That is a reliability problem waiting to happen. So observability, incident response, safe rollout, guardrails, all of it matters more, not less. SRE becomes the adult in the room for a system that is now writing a lot of its own code. The best teams I've seen are investing more here as they lean into agents, not treating it as overhead to cut. Then there's a role most orgs don't have a name for yet. Someone has to own the agent setup itself. The tooling, the context the agents get fed, how the whole thing is orchestrated and evaluated. On strong teams this quietly becomes a full job, usually one of your best engineers, because if they get it wrong it poisons everyone else's output. Some call it an AI labs team. Some fold it into QA, since the evaluation work overlaps. Either way, it's real, and it sets the quality ceiling for the entire org. There's a pattern under all of this. Agents are cheap at producing work and useless at guaranteeing it's right. So value moves from producing to verifying. The roles that felt like cost centers, the ones focused on quality and reliability, are the ones that get more important as the machine takes over the typing. Next in the series: what changes for managers, directors, and VPs.
-
Anil Virmani shared thisEveryone's Asking If AI Agents Will Cut Engineering Jobs. Wrong Question. Everyone asks whether AI agents will shrink engineering teams. That's the wrong thing to worry about. Headcount matters less than what each person is doing, and that's where the real change happens. Start with the junior engineer. The old path was: take small, well-defined tasks, learn by grinding through them, slowly earn bigger ones. Agents are good at exactly those small tasks, so the bottom rung changes first. That doesn't mean you stop hiring juniors. It means the job you hire them into is different from day one. They learn to direct and check an agent's work far earlier than any of us had to. Senior engineers change in the opposite direction. Their value used to be in how much they could build. Now it's in how well they specify what should get built and judge whether what came back is right. With agents handling implementation, they spend far more time up front: understanding requirements deeply, doing the architecture and design, catching the subtle thing an agent got wrong. A senior engineer who can write a spec tight enough to execute and design the system it fits into is worth more than one who just writes code fast. The skill moved up a level, from writing to defining, designing, and verifying. The product manager feels this shift harder than anyone. When building was expensive, the cost itself filtered out mediocre ideas. Agents removed that filter. You can now build the wrong thing faster than ever. So the PM's job stops being about writing more requirements and starts being about killing the ideas that shouldn't get built at all, which means moving closer to the customer and spending more time on the real problem than on documenting the solution. The bottleneck moved from "can we build it" to "should we," and the PMs who get that matter more than ever. Don't read this as everyone becoming a PM. Engineers move up into architecture and technical judgment, but the call on whether something should exist stays a distinct skill and gets more valuable. Cut PMs here and you drown in features nobody asked for, built fast and aimed at the wrong problem. None of this shows up on a boxes-and-lines chart. The boxes look the same. What changed is what happens inside each one, and the orgs that win see it early.
-
Anil Virmani shared thisYour Agentic AI Coding Tools Aren't the Problem. Your Workflow Is. Most teams bolt AI onto the side of how they already work. A coding assistant here, a chatbot there. Then they wonder why the gains never show up in anything the business actually measures. That's the wrong way to think about it. An agent isn't just a faster way to type code. It can take a whole task and run it start to finish while you're not watching. That's a different kind of worker, and if your process still assumes a person at every step, the speed disappears. You've got an agent that can run for hours on its own, stuck inside a workflow that checks in with a human every few minutes. Here's how I've come to think about it. The old SDLC assumed a human at every step. Write the spec, write the code, review it, test it, ship it. Every stage waited on a person. Agents break that assumption. A well-scoped agent can take a spec and produce working code while everyone sleeps. But that only works if the spec is clean enough to hand off, and if you have a way to trust what comes back. So the architecture splits in two. One side is human judgment: what are we building, why, and what does good look like. The other side is agent execution: turning that judgment into code, tests, and fixes at a speed no team could match by hand. The whole thing lives or dies on the specification that connects the two. That connection is where the real engineering work moves to. Not typing code faster, but writing specifications precise enough that an agent can execute them, and building the checks that catch it when the agent gets it wrong. Feed an agent a vague spec and it will hand you back something that looks right and isn't. Feed it a tight one and you get code you can actually ship. Evaluation stops being a QA afterthought and becomes the thing that makes the whole model safe to run. The teams getting this right are not the ones with the most AI tools. They redesigned the workflow around this split, then measured whether the output held up in production. Everyone else is adding assistants to a process that was never built for them and calling it a transformation. That's the gap. The teams that win, treat this as an architecture decision. The ones that stall, treat it as a shopping list of tools.
-
Anil Virmani shared thisThe One Thing Hyperscalers Can't Just Build Themselves A major cloud provider does not lack engineering talent or capital. It can build almost anything it wants, and usually does. So when one of them came to the storage company I was running engineering for and asked us to build a managed service together, that request said more about the state of AI infrastructure than any roadmap slide could. Training large models moves enormous amounts of data in and out of compute clusters, constantly, without room for bottlenecks. Get that wrong and you are paying for idle GPUs while data crawls through the pipe. The hyperscaler had the compute. It did not have the specific storage architecture built for that exact problem, at that exact scale, and building it from scratch would have taken years it didn't have. So we partnered instead. Our high performance file system became a managed offering inside their cloud, sold under their name, running on their infrastructure, built on our engineering. Customers got the speed. The hyperscaler got a capability gap closed fast. We got a new recurring revenue line and a validation that mattered more than the revenue: the world's largest cloud platforms cannot always build their way out of a hard infrastructure problem. The same pattern shows up even faster with the neoclouds, built from the ground up to rent GPU capacity at scale. They move fast and stay lean by design, which means they rarely have the years of engineering depth that storage at this scale demands. For them, partnering with a storage specialist isn't a choice made after hitting a wall. It's in the plan from day one, because standing up a GPU cloud without solving the storage bottleneck first means selling compute that customers can't actually use efficiently. This keeps happening, and it is no coincidence. Storage at AI scale is a narrow, brutally hard problem. Distributed systems that stay fast under sustained pressure, low latency across massive parallel reads and writes, reliability most companies never have to think about. And underneath it all sit the two cardinal sins of storage: never lose data, and never lose data availability. Be fast and cheap and clever all you want, but violate either once and nothing else you built matters. The lesson for anyone building in this space: depth still beats scale. A hundred engineers who have spent a decade solving one hard problem will often out execute a much larger org solving it for the first time. Hyperscalers and neoclouds know this too, which is why they keep showing up as partners instead of competitors. What's changing now is who gets asked to partner. As AI workloads keep growing, the storage layer stops being a commodity purchasing decision and starts being a strategic one, and the vendors who understand that shift early are the ones these providers come looking for.
-
Anil Virmani shared thisWhy I Stopped Asking for Consensus Early in my career I thought a good decision was one everybody agreed with. So I'd keep the discussion open. I'd go around the room again. I'd wait for the last holdout to come around, because a decision without full buy-in felt fragile to me. What I actually built was a team that was slow and quietly resentful, because the people who'd made up their minds early were stuck waiting on a process that treated agreement as the goal. Consensus and commitment are not the same thing, and confusing them costs you more than almost anything else in engineering leadership. Consensus means everyone agrees. Commitment means everyone owns the decision and executes it fully, even the people who argued against it. You can have the second without the first. In fact, the fastest teams I've led usually did. The teams that moved slowest were the ones where nobody would say the hard thing, so nothing ever felt settled enough to act on. The teams that moved fastest argued openly, sometimes heatedly, in the room. Then a decision got made, and everyone got behind it. That only works if two things are true. People trust that their disagreement was actually heard, and people trust that once the call is made, re-litigating it is off the table. I see this constantly now with AI adoption. A team debates which agent framework to standardize on, or whether to trust AI-generated code in a given workflow, and the debate never ends because everyone is quietly waiting for unanimous comfort that will never come. Meanwhile a competitor picked a direction, committed, learned from being wrong in a few places, and pulled ahead. The cost of a slightly wrong decision made fast is almost always lower than the cost of the right decision made too late. So I changed how I run decisions. I make it clear up front whether something is a discussion I'm gathering input on or a decision I'm going to make. I give real space for disagreement, because commitment without genuine debate is just compliance, and compliance falls apart the moment things get hard. Then I make the call, name it clearly, and expect the room to move as one, including the people who lost the argument. And I protect that. When someone tries to quietly reopen a settled decision three weeks later through side channels, that's the thing I shut down, not the original disagreement. The leaders I respect most are not the ones who get everyone to agree. They're the ones who can hold real conflict in a room, close it cleanly, and walk out with a team that's fully committed to a direction half of them didn't pick. That skill matters more now than it ever has, because AI is forcing more decisions, faster, with less certainty than we're used to. Waiting for consensus was always expensive. Now it's fatal.
Anil Virmani commented on a post
7h
Very cool, Michael Dosik !
-
Anil Virmani liked thisAnil Virmani liked thisMy heart was bursting with joy this past weekend as my daughter, Kiera Hua, married her best friend, Curtis Nishihira. ❤️ As my wife, Antonette Hua, and I walked her down the aisle, I was smiling ear to ear while crying tears of joy at the same time. There are moments in life when years of memories are condensed into a few seconds. I thought about meeting Kiera for the first time when we adopted her in Guangzhou, China, at 14 months old. I thought about the little girl we brought home who smiled and laughed all the time, who loved sweets, and who hated the feeling of sand on her feet. I thought about the first time she said, "Baba." ("Dad" in Mandarin.) And then I looked beside me and saw this remarkable young woman beginning a whole new chapter of her life. What a journey it has been. The road has been long and winding, with breathtaking highs and frighteningly dark lows. Through it all, Kiera has become stronger and wiser. Antonette and I are enormously proud of the woman she has become—resilient, kind, determined, joyful, and deeply loving toward her friends and family. This wedding wasn’t simply a celebration of a happy occasion. It was a celebration of who she has become. And somehow, my heart got to swell twice that day. Our son, Stephen Hua, officiated the wedding. His words were thoughtful, funny, wise, and deeply personal. Hearing one of my children speak so beautifully about the other brought tears to my eyes all over again. And then there is Curtis. Curtis, I am so grateful for who you are and the way you love our daughter. You are kind, considerate, adaptable, fun-loving, driven, and a wonderful partner to her. (You also love Star Wars and the MCU, which is a plus.) What I treasure most is that the two of you aren’t simply in love. You’re best friends. You laugh together. You support one another through thick and thin. You bring out the best in each other. (And the fact that you convinced Kiera to take up skiing despite her utter disdain for cold weather is truly a miracle.) That’s a beautiful foundation for a marriage. People sometimes talk about marriage as "giving away" a daughter, but that’s not how it felt to us. We aren’t losing a daughter. We’re gaining a son and another family. Kiera and Curtis now have two families who love them deeply, are cheering for them, and will be there for them through whatever life brings. May your life together be filled with love, laughter, grace, growth, soup, boba, pickleball, and many adventures. I can't wait to see the beautiful future you create together, Mr. and Mrs. Nishihira! ❤️
-
Anil Virmani liked thisAnil Virmani liked thisI often have several Claude Code sessions running at once across a few Macs, and I'd lose track of which one was doing what, let alone how much of my Anthropic quota they were burning through. So I built a dashboard for it in Home Assistant, which I'd only recently started playing with and which turns out to be a great place for this, even if a Claude Code session is a liberal definition of a home device. It's doing one of the things HA is genuinely good at, storing time series data and drawing the trends out of it. Every session on every Mac reports in: what it's working on, which model, how full its context window is, how many tokens it has used. Above that sit two gauges for the five hour and weekly quota windows, and a card that does the arithmetic I actually care about. For example, I can see that 47% of the session's token quota is consumed when only 39% of the time window has gone by which means at the rate I'm burning tokens I'm going to run out before the window resets and refills. I get the same visibility for the weekly quota as well. And the code is open source, which is a first for me. https://lnkd.in/gkZPWfj2Using Home Assistant to Herd a Gaggle of Claude SessionsUsing Home Assistant to Herd a Gaggle of Claude Sessions
-
Anil Virmani liked thisAnil Virmani liked thisAfter an unforgettable journey of 16+ years across two tenures, building, scaling, and operating some of the most complex infrastructure and networks in the world, the time has come for me to send my final FIN-ACK and initiate a graceful handoff at Google. Looking back at where we started and what we have accomplished together with the #GoogleGlobalInfrastructure team, I am filled with immense pride and gratitude. From laying the groundwork for a resilient global fabric, to pushing the envelope on modern consumable infrastructure, none of it would have been possible without the relentless curiosity, execution rigor, and unwavering commitment of the Google infrastructure teams to keeping the packets forwarding, the tokens serving, and the queries flowing. Beyond the technology, product, designs, and scale, what I will cherish most is the people. It has been an absolute privilege to work alongside such brilliant minds, solve impossible problems during late-night war rooms, celebrate our shared milestones, and always strive to build the infrastructure sustainably and the right way. There are countless colleagues, mentors and leaders to thank. I am deeply grateful for the generosity shown to me through every high and low by Urs Hölzle, Thomas Kurian, Amin Vahdat, and Ruth Porat—thank you all from the bottom of my heart. I’ll be joining the Blackstone | Google joint venture, Crux AI, as the Chief Technology & Product Officer. The pursuit to engineer and scale next-generation infrastructure continues, and I am thrilled to partner once again with Benjamin Treynor Sloss to tackle what comes next.
Experience & Education
Licenses & Certifications
Languages
-
English
-
-
Hindi
-
Recommendations received
4 people have recommended Anil
Join now to viewOther similar profiles
-
Vijay Gill
Vijay Gill
Great engineering organizations as a product: Culture, recruiting, leadership, execution and communications architecture.<br>My job is to make everyone in the engineering organization successful and unblock issues that are getting in the way of success of the team.<br><br>Our mission: Empower Developers Around the World to Create Transformative Software Applications
-
Prince Kohli
Prince Kohli
Explore more posts
Explore top content on LinkedIn
Find curated posts and insights for relevant topics all in one place.
View top content