Nicky Pike

Field CTO at Coder and host of the [Dev]olution podcast

Shadow IT was the polite version.

We used to worry about a developer spinning up a rogue cloud account or sneaking some unsanctioned SaaS past the security team. Annoying, sure. But it moved at human speed, and you could mostly catch up to it.

Gari Singh has a better name for what&#39;s coming next. He calls it Ninja IT in the latest episode of <a href="https://youtube.com/@Devolution-Podcast/featured">the [Dev]olution Podcast</a>.

Coding is Using Tech to Solve Problems

Gari&#39;s a product manager at Google Cloud working on the containers and Kubernetes runtimes carrying a big chunk of the world&#39;s AI workloads, and before that he put in 15 years at IBM as a distinguished engineer and a blockchain CTO. So he&#39;s watched three infrastructure waves wash through the enterprise. When he describes the current one, it&#39;s not agents quietly bending the rules. It&#39;s agents executing in secret, in their own little groups, wiring themselves to systems at a speed no human approval chain was built to keep up with. (<a href="https://youtu.be/qTRwwcx4QQI?t=1150">He coins it around 19:10.</a>) Shadow IT at least waited for a person. This doesn&#39;t.

And here&#39;s the part that should make a platform engineer sit up. <a href="https://www.deloitte.com/us/en/about/press-room/state-of-ai-report-2026.html">Deloitte&#39;s 2026 State of AI survey</a> found that close to three-quarters of companies plan to deploy agentic AI within two years, but only 21% have a mature model for governing it. So the stampede is already in motion, and four out of five companies are running into it without the controls to hold the line.

Let&#39;s be honest about what makes this hard. An agent doesn&#39;t sit there waiting for instructions like a command line that runs once and dies. It runs in a loop, calling the model, picking tools, acting on what comes back, going again. Point it at a problem and it&#39;ll reach for every tool in the box to solve it. If Bash is on the machine, it&#39;ll use Bash. Nobody told it not to.

So what does a platform engineer actually do about it on Monday morning? That&#39;s the back half of the latest episode of the [Dev]olution Podcast, and Gari&#39;s answer is refreshingly un-fancy. It isn&#39;t &quot;buy a firewall&quot; and it isn&#39;t &quot;ban the agents.&quot; It&#39;s three specific things you can start this week, and the third one is the one that&#39;ll get under your skin: go build one of these things yourself and watch what it does when you&#39;re not looking. (<a href="https://youtu.be/qTRwwcx4QQI?t=2651">He walks through the checklist around 44:11.</a>) Also on the table: why he thinks MCP servers are the new SOAP, not the new HTTP, and what a no-op user mapping in a container has to do with surviving the next breach.

<a href="https://youtu.be/qTRwwcx4QQI">Watch the full episode on YouTube</a>

If agent governance, infrastructure at machine speed, and a Google PM cheerfully telling you to go scare yourself sound like your kind of conversation, subscribe. New episode every two weeks.

Shadow IT was the polite version. Gari Singh has a better name for what agents are about to do to your stack, and platform teams won't like it.


Ninja IT: Gari Singh on the AI Infrastructure Nobody Invited the Platform Team To

Peter Davies

Enterprise Account Manager, International Government

Ukraine ran a war on commercial cloud infrastructure. Not because it was the ideal setup, but because Russian strikes had made physical data centers a liability, not a safeguard. The mission continued anyway.

That single fact is quietly rewriting how defense and government organizations define sovereignty. For decades, sovereignty meant hardware on national soil, physical custody of servers, and air-gapped isolation from anything resembling the open internet. That model was built for a static threat environment. It is being tested against a dynamic one, and it is losing.

The organizations rethinking this fastest are the ones already under the most pressure: defense ministries, prime contractors, and government agencies racing to adopt AI without losing control of classified code and data. What they are converging on is a different definition of sovereignty, one built around resilient access and enforced control rather than geography.

<h2>The old sovereignty model was built for a world that no longer exists</h2>

Physical isolation made sense when the primary threat was a foreign government gaining direct access to a server rack. It makes less sense when the primary threat is an AI agent with too much network access, a compromised credential, or a supply chain dependency nobody audited.

Defense organizations are not abandoning control. They are relocating it. The real question, as one HPE executive put it to a room of NATO officials at TechNet, is not where the data center sits. It is who controls access and who can see it. Those two answers define an organization&#39;s actual security posture, not the address on the building.

The market is already voting on this. <a href="https://www.fortunebusinessinsights.com/sovereign-cloud-market-112386">The latest market analysis</a> pinned the global sovereign cloud market at $154.69 billion in 2025 and is projected to reach $195.35 billion in 2026. That growth is driven largely by government and public sector demand for infrastructure that offers control without requiring physical isolation.

<h2>AI adoption in government has moved past the pilot stage</h2>

For the last several years, government AI initiatives lived mostly in pilot programs and press releases. That phase is over. <a href="https://www.brookings.edu/articles/from-governance-to-execution-in-federal-ai-policy/">Brookings</a> found that the value of federal AI funds obligated increased 966% between 2024 and 2026, rising from $355 million to $7.2 billion, with the Department of Defense accounting for $90 billion in AI contract spending alone.

The number of federal agencies holding AI contracts rose from 17 in 2022 to 28 in 2026, out of 441 total agencies. That is not experimentation. That is procurement at scale, moving faster than most agencies&#39; existing security review processes were built to handle.

Speed at this scale creates a new kind of exposure. Every agency now racing to deploy AI coding tools and autonomous agents is also racing past the infrastructure questions those tools require: where do agents run, what can they touch, and who is watching when something goes wrong.

<h2>Governance is what makes speed sustainable, not what slows it down</h2>

The instinct in a fast-moving procurement environment is to treat governance as friction. The data says the opposite is true. According to <a href="https://coder.com/blog/what-100-engineering-teams-revealed-about-ai-maturity-and-what-to-do-about-it">Coder&#39;s 2026 AI Maturity Assessment</a> of 100 engineering organizations, 78% are already running AI agents in some form, but 69% have no security protections or only ad-hoc controls against data exposure and secret leakage.

That gap between adoption and control is where the real exposure lives: agents with broad access to code and systems, operating without the safeguards that would catch a leak or contain a mistake. In a defense context, that gap is not a productivity annoyance. It is classified code and sensitive data moving through infrastructure nobody is actively protecting.

The alternative is not slower AI adoption. It is AI adoption that is legible from the start, with humans in the loop and every agent action attributable to a person and a policy. Organizations that build this in from the beginning move faster later, because they are not stopping every six months to retrofit controls onto systems that were never designed for them.

<h2>What this looks like inside a classified environment</h2>

One U.S. defense intelligence organization felt this gap directly. Engineers spent days manually configuring local machines before they could write a line of code, and every setup drifted slightly from the last, turning routine debugging into environment archaeology. Leadership had no centralized way to see who was working on what or whether development practices matched security policy.

The organization moved development onto <a href="https://coder.com/success-stories/defense-intelligence-organizations">Coder, self-hosted on AWS GovCloud inside a fully air-gapped network</a>, with every environment defined as a Terraform template rather than a manually configured machine. Onboarding time dropped from multiple days to three to five minutes, and usage grew 500% over the following years as engineers pulled the platform into more teams once they saw it work.

The pattern holds at a larger scale, too. A separate <a href="https://coder.com/success-stories/u-s-defense-intelligence-organization">U.S. Defense Intelligence platform</a> built the U.S. military&#39;s first true multi-tenant Coder deployment, giving more than 2,500 developers pre-approved, ATO-compliant environments instead of forcing each 20-person mission team to certify its own infrastructure from scratch. Centralizing that approval process did not just speed up onboarding. It gave the organization one place to enforce the same security posture across every mission team, at any scale.

<h2>What comes next is a test of execution, not intent</h2>

Every argument above resolves at the same layer: where code actually gets built and what it can touch while it is being built. Commercial enterprises can get away with &quot;build it on a laptop and figure out governance later.&quot; Classified and regulated environments cannot.

Coder&#39;s <a href="https://coder.com/solutions/ai-governance">AI governance</a> capabilities exist for exactly this reason: agents and developers work inside self-hosted, policy-controlled infrastructure, with every model call authenticated and logged and every agent&#39;s network access restricted by default rather than by exception. Coder Agents extend that same model to autonomous coding agents specifically, running on infrastructure the organization already controls, whether that is a cloud VPC or a fully <a href="https://coder.com/solutions/agents">air-gapped environment</a>, with no API keys scattered across individual workspaces. A defense agency evaluating AI adoption today is really evaluating one question: if an agent misbehaves, what is the blast radius, and who finds out first.

The demand signal here is public and it is not subtle. The EU&#39;s <a href="https://defence-industry-space.ec.europa.eu/eu-defence-industry/white-paper-european-defence-readiness-2030_en">Defense Readiness 2030 agenda</a>, NATO&#39;s <a href="https://www.nato.int/en/about-us/official-texts-and-resources/official-texts/2025/06/25/summary-of-natos-rapid-adoption-action-plan">push to compress procurement cycles</a> from years to months, and the sharp rise in <a href="https://www.brookings.edu/articles/from-governance-to-execution-in-federal-ai-policy/">U.S. federal AI contract spending</a> all point in the same direction. The organizations that win are not the ones with the boldest AI strategy on paper. They are the ones that already built the governed infrastructure to execute it before the mandate arrived.

Coder <a href="https://coder.com/blog/coder-secures-and-accelerates-mission-critical-innovation-in-public-sector">works with public sector and defense organizations</a> building exactly that kind of infrastructure today, on their own terms and their own hardware. If your teams are already navigating this shift, an <a href="https://coder.com/solutions/ai-governance">AI governance assessment</a> is a reasonable next step before the procurement cycle forces the question.

Air-gapped servers no longer equal sovereignty. See why AI agent security, not geography, now defines control across defense infrastructure worldwide.

Why Sovereignty in Defense No Longer Means Air-Gapped Infrastructure

Everyone&#39;s still arguing about whether AI writes good code. Gene Kim thinks that&#39;s the wrong fight.

I went into this episode of the <a href="http://youtube.com/@Devolution-Podcast/featured">[Dev]olution Podcast</a> with one question. Gene spent 27 years measuring what makes great technology organizations just so great, which led to the work behind the DORA metrics and The Phoenix Project, so I wanted his read on the thing nobody&#39;s really asking about vibe coding: when anyone can build software, what breaks first?

He didn&#39;t hedge. He did the math live.

Nobody Knows Until Teams Start Trying

<h2>The math, out loud</h2>

Twenty-eight million developers on the planet today. Multiply by a hundred, the multiplier Gene thinks this shift brings, and you land at 2.8 billion people building software. A third of the planet. It sounds absurd until you remember Gartner already expects <a href="https://venturebeat.com/business/gartner-citizen-developers-will-soon-outnumber-professional-coders-4-to-1">citizen developers to outnumber professional coders four to one</a> inside big companies. Gene just took that curve and kept drawing it: the ER doctor, the product owner who never wrote a line, the 35-year police veteran, all of them shipping now.

<h2>Where vibe coding actually breaks</h2>

Here&#39;s the surprise, and it&#39;s the answer to my question. The constraint doesn&#39;t vanish when coding gets cheap. It moves. Gene&#39;s image for it stuck with me: in the vibe coding book he wrote with Steve Yegge, he says it&#39;s like <a href="https://youtu.be/xOwsHtahx2g?t=1566">Disneyland when every ride becomes instant</a>. Everybody&#39;s still standing in line, and your whole job as a leader is making sure nobody&#39;s waiting.

You can already see where the line forms. Faros AI pulled telemetry from more than 10,000 developers across 1,255 teams, and the teams leaning hardest on AI <a href="https://www.faros.ai/blog/ai-software-engineering">merged 98% more pull requests while review time climbed 91%</a>. More code, faster, stacking up in front of the step that still comes down to human judgment: how much oversight it needs, and how fast that trust gets earned. That same study found none of the speed reached company-level delivery metrics, so the individual gains are real and the org-level payoff is still unproven.

<h2>The part I didn&#39;t expect</h2>

I&#39;ve made a version of Gene&#39;s argument myself. I compared AI to Home Depot: it put building within reach of people who never got the license for it. That comment lit the usual fire on LinkedIn with the &quot;vibe coders should never touch code&quot; crowd. Some of what comes out is slop, no argument. But writing off the whole thing misses what&#39;s new, which is that the barrier just fell from years of hard-won skill to a cost almost anyone can cover, for the first time.

What I didn&#39;t expect from a man with 27 years of measurement behind him was how comfortable he is not knowing yet. Ask Gene the biggest open question and he answers, &quot;where do I start.&quot; He rebuilt a legacy project overnight in a single shot, woke up to working software, and admitted that by his own old rules he&#39;d done it all wrong. He <a href="https://youtu.be/xOwsHtahx2g?t=1965">called it magic anyway</a>. He&#39;s even questioning whether 25 years of building on open source is still an asset or has quietly become a liability.

I won&#39;t tell you where he lands on that, mostly because he isn&#39;t sure either, and that&#39;s the honest part of this whole moment. The measurement guy, comfortable saying nobody has the answer yet.

<a href="https://youtu.be/xOwsHtahx2g">Watch the full episode on YouTube</a>

If a 2.8-billion-developer future, the researcher behind DORA questioning his own gospel, and an honest &quot;we don&#39;t know yet&quot; sound like your kind of conversation, subscribe. New episode every two weeks.

Gene Kim on vibe coding: he did the math live, 2.8 billion builders. What breaks first isn't the code. It's the line they're all standing in.

When Anyone Can Build: Gene Kim on What Vibe Coding Breaks First

Discover the latest at the official Coder blog: updates, company insights & AI development infrastructure tips for enhanced developer productivity.

Ninja IT: Gari Singh on the AI Infrastructure Nobody Invited the Platform Team To

Why Sovereignty in Defense No Longer Means Air-Gapped Infrastructure

When Anyone Can Build: Gene Kim on What Vibe Coding Breaks First

All posts