Coté

The New Bottleneck That Will Chill Your AI Vibes

Thu, 02 Apr 2026 13:47:42 +0200

Platforms are helpful, from 2025 DORA Report.

Getting software from idea to production quickly has always been important, but the increase in AI-driven application deploys is about to overwhelm even the best laid golden path to production. We can feel in our bones developers are already revving apps more frequently. The 2025 DORA Report found that 95% of developers use AI and over 80% say it has made them more productive. And that is just developers.

Now imagine every employee in a company - marketing, operations, finance, HR - using agentic AI to build and deploy their own applications. Not as a one-off experiment, but iterating multiple times a week, even daily. Gartner thinks that 80% of technology products and services will be built by people outside traditional IT by 2026.

"AI adoption now improves software delivery throughput... However, it still increases delivery instability. This suggests that while teams are adapting for speed, their underlying systems have not yet evolved to safely manage AI-accelerated development." - 2025 DORA Report

The volume of software hitting your infrastructure is about to multiply in ways that make the current CI/CD pipeline look like memories from a long ago golden era. Without a platform to handle the builds, the deployments, the security scanning, the compliance checks, infrastructure becomes the bottleneck that stymies enterprise dreams of AI productivity gains and ROI. I think that is part of why, as of Spring 2026, 56% of companies report getting very little out of AI.

Without a real application platform and a supporting platform engineering team, the benefits you get from AI will be negligible. Everything will bottleneck at the deployment and runtime stage when your systems are overwhelmed by the velocity of AI apps. With a platform, that bottleneck can be eliminated.

A good platform turns the flood of AI-generated applications into a manageable flow. An AI-minded platform also adds in all those enterprise-y features like security controls, data access, and the runtime management needed to keep those apps running and performant. You know: "day two."

Without a platform, getting a new app to production means a ticket to the infrastructure team, a security review, weeks if not months of meetings. As has been the case for over a decade, a real platform replaces all of that with a self-service golden path.

When the number of people who can create software expands from your development team to your entire workforce, what a platform gives you goes from "nice to have" to "the thing that determines whether AI investment pays off or just creates more enterprise sludge."

So... unless you are lucky enough to already have it, this is a pretty good time to TryTanzu.ai.

Thu, 02 Apr 2026 09:13:59 +0200

Check out the live reacording of this week’s Software Defined Talk. Final episode posted tomorrow, on Friday if you prefer traditional podcast.

Histoires Prodigieuses, Forward Deployed Vibes, and Skyrgamur - Related to your interests, Wednesday

Thu, 02 Apr 2026 07:41:03 +0200

Also: Spring still number one, KubeCon AI workloads, the end of Doctor No, and a buyer’s market for employees.

Spring still king, from The State of Java 2025, JetBrains.

🤖 From Red Hat OpenShift to VMware Cloud Foundation 9: A Journey to a Unified Cloud Operating Model - A defense-industry case study describing a migration from OpenShift running on vSphere to VMware’s native Kubernetes stack (VKS + VCFA + VCF 9), framed around operational consolidation and TCO reduction.
Will MS Copilot Cowork Enable Real Enterprise AI Collaboration? - “Individual AI copilots are well understood at this point. What remains genuinely hard is maintaining a coherent AI context across multiple participants with different roles, permissions, and intentions in a live session.”
Here’s what that Claude Code source leak reveals about Anthropic’s plans
Block the Prompt, Not the Work: The End of “Doctor No” - “They had blocked the website. They hadn’t blocked the risk.”
The State of Java 2025 - Spring usage still number one.

AI workloads in Kubernetes.

Six Takeaways From KubeCon EU 2026 - Good round-up of topics at KubeCon, especially using Kubernetes to run AI workloads.
Almost three-quarters of US workers think now’s a bad time to find a quality job - Buyer’s market for employees.
Elizabeth I’s Manuscript Copy of Pierre Boaistuau’s Histoires Prodigieuses (1559) - So many fun-as-in-weird images.
Book review: Maintenance: Of Everything (Part One) - This book review feels like it could be about most business book reviews, and wherever genre the book in question is. Namely: “This is fascinating. Tell me how? Tell me what was needed to make it happen? But, unfortunately, outside of some basic tenets of ‘give the rank and file more freedom to do things’ and ‘embrace improvisation,’ the book doesn’t seem to offer more.”

Wastebook

Skyrgamur - when in Iceland during Christmas, keep an eye on yogurt.
Snark! ‘Quote: “Our favorite holding period is forever”. Action: Trade Apple to realize gain.’ Here.
“Forward Deployed Vibes.” Here.
“…franchise…lettuce…offerings…” Kenji.
“the nice thing about picking your own battles is that when you’re battling yourself, you can let yourself win by not caring too much.” Ibid.

Logoff

See you next time!

Five Disneys, Conviction Collapse, and sk8erboi2006 - Related to your interests, Tuesday

Wed, 01 Apr 2026 20:35:04 +0200

Also: local models worth running, Susan Sontag on list-making, GraalVM as thriller novel, and the slow ROI of children.

The Prodigal Daughter (1903), John Collier. Via New Cartographies.

Recommended local models - “I’ve experimented with the following models and have found useful applications (mostly around chat, MCP and coding) with all of them: gpt-oss-120b; Qwen3-Coder-30B; GLM-4.7-Flash; MiniMax-M2.5; Qwen3.5-35B-A3B”
On Critical Thinking in a Multi-Model World - Getting the AI to be helpfully disagreeable, thus, engaging and truth-finding rather than satisfying.
Ollama Now Runs Faster on Macs Thanks to Apple’s MLX Framework - “the new version processes prompts around 1.6 times faster” and better memory management.
“This is unprecedented”: America’s AI boom is leaving the rest of the world behind
OpenAI raises $122 billion to accelerate the next phase of AI
Catching Up with GraalVM: When “Faster Java” Turns into Patch Tuesday - and a Clean Break with Intel Macs - This is one of the most action-oriented articles about Java native compilation you’ll ever read. It’s a few kidnappings and murders away from being a gripping thriller novel.
YAML is (not) my preferred configuration format
HolmesGPT
Oracle believed to have cut 10,000 positions across multiple divisions as mass layoffs begin to fuel AI investments
Susan Sontag’s Playground of Ideas - “I perceive value, I confer value, I create value, I even create - or guarantee - existence. Hence, my compulsion to make ‘lists.'” / I’m not writing it down to remember it later, I’m writing it down so it exists.

Susan Sontag's papers, "Death-Kit Notes," UCLA, Box 125, Folder 7. Via Jillian Hess.

Wastebook

“trad agency model” here.
“a ~$7 billion-a-year spice and sauce giant” flavor.
“That’s where we’re at, when joy means paying your bills.” Dickerson
“They fit into the margins of the day in a way nothing else could.” They are podcasts.
The slow ROI of children: “Then one day they can fetch things.” 7 Reasons Kids Are Overrated Until Suddenly They’re Your Entire Support Infrastructure
“part of the deal, French films were only allowed to be shown for four weeks out of thirteen.” Read to the end for a nuclear joke. Naughton
“maybe 13-year-old you signed up for sk8erboi2006@gmail.com, but as an adult you would prefer to have a more professional johnsmith@gmail.com address.” Don’t Like Your Gmail Email Address? You Can Finally Change It
“…about five Disneys.” OpenAI is now valued at $852 billion
“conviction collapse.” “Conviction Collapse” and the End of Software as We Know It

ICYMI

Using HATEOAS for agentic AI - a demo
VMware Kubernetes Service (VKS) architectural diagram
This cake is bananas
Field Engineering is the YOLO team - SOFTWARE DEFINED TALK #565 - “This week, we discuss big bets gone wrong, OpenAI buying the Python toolchain, and Claude’s favorite tools. Plus, how fast should AI agents ship bug fixes?”

Logoff

See you next time!

Using HATEOAS for agentic AI - a demo

Tue, 31 Mar 2026 12:18:02 +0200

My co-worker Adib did a video showing how to use Spring HATEOAS to make existing REST APIs agent-friendly. Instead of wrapping your 700 APIs as MCP servers and dumping all those tool definitions into the context window, you build a an adapter layer that lets agents discover what they can do by following links in the spirit of the World Wide Web REST thinking from long ago: HATEOAS.

Progressive disclosure

This method is geared around minimizing context window usage for especially large APIs. As the AI follows links, it learns more. For example, if you have the API to get a person’s bank accounts, it can then “follow the link” to a bank account to get more info - rather, to get the API and info for getting the account details. Then it can follow the link to the ledger, getting the API and details. Then the individual transactions. And so, and so on. Like skills, it’s progressively discovering what the APIs are and preserving the context window.

Permissions

This also give you more power over access and permissions. On the server side, you can determine what API and datas you expose to AIs walking over the data. You could turn off transaction details, you could mask the edit being sent back to remove PII, etc.

Just turn it on

The other nice thing is that HATEOAS and REST are very well understood and, thus, easy to implement. As Adib shows, you can kind of just unleash a standard HATEOAS pattern or library onto an API and have it map it, like Spring HATEOAS. This gives you another layer of control as well. It also means you can use standard HTTP auth, and I assume all the day 2 stuff like load balancers, caching, etc. It’s all just web traffic to that stuff.

Anyhow, it’s a cool idea. Check out the video.

VMware Kubernetes Service (VKS) architectural diagram

Mon, 30 Mar 2026 14:01:55 +0200

At KubeCon EU, March 26th, 2026.

This cake is bananas

Mon, 30 Mar 2026 13:29:23 +0200

Hot modems, dinergoths, the platform bottleneck, and the SaaSpocalypse - Related to your interests, Monday

Mon, 30 Mar 2026 07:34:02 +0200

Also: sovereignty’s control plane, 81,000 people tell Anthropic what they want, Google sells its fiber, and oats for the sparrows, and cake

My pal Adib Saikali wrote up an MCP security guide covering how to think about securing MCP servers in the enterprise (no lead-generation required, just a straight-up PDF download). It gets into access tiers (open, group, and user-level servers), authentication with OAuth 2.1, identity propagation models (when to use service accounts vs. forwarding user identity), and how an MCP gateway gives you a governed chokepoint for auth, observability, and capability filtering. Good starting point if you’re deploying MCP servers that touch real corporate data and don’t want your AI agent behaving like a rogue insider. Check it out.
Enterprise dev teams are about to hit a wall. And CI pipelines can’t save them. - The platform is now the bottleneck, again. // “The math is straightforward. If each change requires 30 minutes of validation in a shared staging environment and an agent-assisted developer generates 5 or 6 PRs a day, the developer spends the majority of their time managing a deployment queue rather than building software. The agent accelerates code output velocity, but if the surrounding system stays slow, that velocity hits a wall.”
The agentic AI gap: Vendors sprint, enterprises crawl - “today firms spend approximately 4% of their revenue on technology. We believe this figure will rise to 10% or more within the next decade. " // And a whole lot of enterprise AI analysis.
the diginomica network pulse - what CIOs actually think about AI and the SaaSpocalypse - “Over a quarter (38%) of the community members who took part in the pulse survey see AI as a direct replacement or competitive threat to existing SaaS technologies in the enterprise. That could lead to a significant disruption to the SaaS marketplace over the next 12-24 months. "
Infrastructure is all that’s left - When the AI is the UI: “Everything that has been build around the centrality of a UI will need to be rebuilt around APIs. I am not saying UIs will disappear over night, but it becomes clear that it is no longer the place where one can accrue a lot of value.”
Cloud-native open-source infrastructure meets AI’s big test - “When I have talked with companies who are creating specialized models, they’re seeing that they’re way faster, way more accurate and way cheaper,” he said. “If you can increase efficiency, then you can really get to that ROI and you can bring forward AI adoption.”
From Building Platforms to Delivering Capabilities: KubeCon + PlatEngDay EU 2026 Summary - Extensive overview of the platform engineering story and content at KubeCon EU.
The Control Plane: The Next Frontier of Infrastructure Sovereignty - Here’s all the governance and management functions it takes to make something enterprise-y.
The 12x Bet on AI - Based of debt raises, he models out that “AI revenue will grow 5x within five years.”
What 81,000 people want from AI \ Anthropic

Matt Eastwood, et. al., IDC.

‘Astound’ed. Google Flips Its Fiber To PE. - “The playbook is not complicated. Low introductory pricing, step-function rate increases after twelve months, retention discounts for customers who call to cancel, and underinvestment in everything that doesn’t show up on a cash flow statement. The BBB and Trustpilot reviews are a graveyard of billing complaints and broken promises.”
How Not to Interview (Interesting People) - 🤖: The five cardinal sins: finishing their sentences, volunteering your own references, mentioning your other work, asking only comfortable questions, and filling silence.
Google Has a Secret Reference Desk. Here’s How to Use It. - Arcane Google search incantations.
Explore - Banner Depot 2000 - “Hot deals on fast modems.”

The Kids These Days

That’s why complacency is key to understanding this framing of downward mobility in 21st century America. The upper middle class needs to believe that the economic woes of others are due to a moral failing. That if the dinergoths really wanted to succeed in America, they could. To contemplate that maybe there are structural forces at play which prevent everyone from being upper middle class is to admit that maybe forces beyond your control could one day push you down that same slope. From 8ball.

Wastebook

“If you’re wondering how cold fusion produces pollution, let alone a lot of it, don’t worry, this is explained in a chapter by one teenaged prodigy tennis player to another, wherein the latter is wearing a blindfold to improve his other senses but can’t find a bathroom, so he asks the other player for help, but instead, the other player explains cold fusion in exhaustive detail over such a long period of time that the blindfolded player is nearly peeing in his pants, and it all turns out to be a ploy by the explainer to pressure the blindfolded player (who is a Muslim and drug-free) to give away his urine so the other player and his friends can pass a drug test. Stuff like this happens a lot in IJ.” Infinite Jest Extraction
Wow, the eye tracking AI thing in Descript actually works. Now we can all read scripts and pretend like we know how to memorize things!
“He made enough money to retire to a farm in Vermont in search of humility.” PE Farm.
“your half-zip won’t save you.” Here.
“If you feed enough oats to the horse, some will pass through to feed the sparrows.” John Kenneth Galbraith on the dogma of ‘trickle-down economics.
“Walking means you can’t fast forward through anything” Getting around Japan.
“vibe porting” Saved $500K/Year

Nicolas Biddy, 1983(?)

ICYMI

Securing AI - More “how do we secure this AI stuff” talk with David Zendzian on today’s live stream. He’s recently gone Claude Crazy so I wanted to get his CISO-supremo talk on thinking through the risk management for AI in enterprises.
Field Engineering is the YOLO team - Software Defined Talk #565 - “This week, we discuss big bets gone wrong, OpenAI buying the Python toolchain, and Claude’s favorite tools. Plus, how fast should AI agents ship bug fixes?”

Logoff

Last week was Mila’s sixth birthday, lots of birthday-fun and parties. To a six year old, every day is a chance for awesome, especially when you get presents, four hours to play with friends, and rainbow cakes.

I think cutting over to micro.blog to post the newsletter is working…right?

Securing AI

Fri, 27 Mar 2026 17:58:18 +0200

More “how do we secure this AI stuff” talk with David Zendzian on today’s live stream. He’s recently gone Claude Crazy so I wanted to get his CISO-supremo talk on thinking through the risk management for AI in enterprises. Each time I tried to come up with a problem, he was good finding the fix. Plus, we talk about some of the things we’ve learned about using our little robot buddies.

And, if you want more, we started talking about security and AI in the episode a few weeks back.

Fri, 27 Mar 2026 08:11:54 +0200

If you’re wondering how cold fusion produces pollution, let alone a lot of it, don’t worry, this is explained in a chapter by one teenaged prodigy tennis player to another, wherein the latter is wearing a blindfold to improve his other senses but can’t find a bathroom, so he asks the other player for help, but instead, the other player explains cold fusion in exhaustive detail over such a long period of time that the blindfolded player is nearly peeing in his pants, and it all turns out to be a ploy by the explainer to pressure the blindfolded player (who is a Muslim and drug-free) to give away his urine so the other player and his friends can pass a drug test. Stuff like this happens a lot in IJ. Infinite Jest Extraction

Rainbow Cake

Thu, 26 Mar 2026 12:13:36 +0200

Mila's sixth birthday. Cake from De Drie Graefjes Zuid.

MCP Security Guide

Thu, 26 Mar 2026 12:02:48 +0200

My pal Adib Saikali wrote up an MCP security guide covering how to think about securing MCP servers in the enterprise (no lead-generation required, just a straight-up PDF download). It gets into access tiers (open, group, and user-level servers), authentication with OAuth 2.1, identity propagation models (when to use service accounts vs. forwarding user identity), and how an MCP gateway gives you a governed chokepoint for auth, observability, and capability filtering. Good starting point if you’re deploying MCP servers that touch real corporate data and don’t want your AI agent behaving like a rogue insider. Check it out.

KubeCon EU 2026, bottlenecks, Corporate Bullshit Receptivity, and the End of the Meat-Mouse - Related to your interests, Wednesday

Thu, 26 Mar 2026 11:33:36 +0200

Also: sovereign AI, Claude Code auto mode, FCC bans foreign routers, Walmart digital price tags

KubeCon EU 2026

KubeCon is crazy big. James says something like 13,500+ people. They’re going for another infrastructure grab: all the stuff used to run AI. It’s worth asking who they’re competing with there. I have no idea how OpenAI and Claude run the AI PaaSes they’v built - the ones that everyone is bonkers for. Do they run on Kubernetes?

If not, that’s the competition. Pay attention to them showing up - or not - at future KubeCons.

On the topic of platform engineering, here’s some quick thoughts on platforms as brought up in another great keynote from Abby Bangser. Preview: not that she’d say it, but Kubernetes is the bottleneck. Throw in how fast developers can move with AI, and that becomes even more clearer than ever.

Claude Crazy

There’s a little collection of us who’ve started using Claude code for real life and knowledge worker stuff - that is, not for programming. I’ve talked about this on our podcast, there’s a great dive into it on Reasoning this week (audio or video), and here’s a crack at me typing it up.

I know, I know: starry eyed and all that. The thing with this is that it’s not about replacing programmers, it’s about removing the bullshit that blocks you from getting stuff done. Listen to the end of the Reasoning episode for an explanation of that.

Now, pardon me as I got off to get some shit done.

Still from They Live.

Broadcom donates Velero to CNCF - and it could reshape how Kubernetes users handle backup and disaster recovery - “During a briefing with the press before the announcement, Broadcom VMware representatives sought to reassure the community that it is ramping up open-source support and contributions - especially for Kubernetes, the largest CNCF project - using KubeCon + CloudNativeCon Europe as the venue to make these announcements. "
Here’s more on VMware at KubeCon.
GKE and OSS innovation at KubeCon EU 2026 - Round-up of Google stuff from KubeCon.
Oracle Embraces AI At JavaOne - “Oracle announced that it has restarted Project Detroit, which should lead to closer integration with foreign AI libraries and tools, starting with JavaScript using V8 and Python using CPython. In concert with Project Panama, this will give Java the ability to access AI technologies written in languages outside the Java ecosystem. After Project Detroit ships, expect to see lots of Java wrappers around the cool libraries that developers use in other environments.” And: “Oracle will need to convince other corporate stakeholders with case studies and by its own example that it’s possible and safe to integrate AI into an enterprise Java application in a way that provides business value.”
Getting real about agentic AI projects - what’s working, and what’s not? - A round-up of elusive enterprise AI ROI.
FinOps expands focus to ROI, AI efficiency in cloud era - “Generative AI accounts for more than half of public cloud services used by enterprises, with large companies investing heavily in governance and oversight.” That from the Flexera 2026 cloud report. Seems insanely high. Did the half of workloads that replaces just disappear? // Indeed, I think this is more about the number of survey respondents saying they use AI (but not how much they spend). From the Flexera survey: “GenAI has become pervasive. In terms of public cloud services used by all organizations, GenAI has jumped 8 percentage points to the third-place spot (58%), and nearly half say they use it extensively.”
Auto mode for Claude Code - “I remain unconvinced by prompt injection protections that rely on AI, since they’re non-deterministic by nature.”
Developer AI Tooling in 2026: Trends Shaping How We Build
Why aren’t AI productivity gains higher?
Datadog bets DIY AI will mean it dodges the SaaSpocalypse - “for AI systems to win trust, their output must be both explainable and verifiable. Using its own models makes that easier for Datadog, she said. They have also helped the company to create a tool that watches AI platforms while they work and can detect signs they are producing hallucinated output.”
Edera spent years calling KVM less secure. Here’s why it changed its mind. - “That investment deserves to be met, not worked around. Edera should work within that architecture. This summer, it will.”
The Control Plane: The Next Frontier of Infrastructure Sovereignty - Here’s all the governance and management functions it takes to make something enterprise-y.
Why sovereign AI is now a top enterprise priority
I’m OK being left behind, thanks! - “Why should I invest in learning the equivalent of WordStar for DOS when Google Docs is coming any-day-now?” This is a fine strategy. Do what works for you, smoke ‘em if you got ‘em, etc.
The Secretary of War didn’t really mean it, contends US government lawyer as Anthropic gets its first day in court - Walk-back in progress.
US FCC blocks new foreign-made consumer routers - Wow, that is a huge move. Better go buy whatever gear you want now.
iOS 26.4 Features: Everything New in iOS 26.4 - Many small things, including new emojis and Music features.
Walmart digital price tags are coming to every US store by end of 2026 - When your customers don’t want digital transformation. // Also, the outrage reminds me of the anti-capitalist/WTO funk of the 90’s. What a simpler time it was when you could express your rage by throwing a brick through a Starbucks window and call it a day.
The Corporate Bullshit Receptivity Scale: Development, validation, and associations with workplace outcomes - the paper. Related: “We need to find some path back to trusting that individuals will do their jobs, without every responsibility being visible to an entire organisation, without follow-ups being scheduled by a cadre of overpaid managers with their overfed metrics.”
Explore The Marginal Revolution Generative Book by Tyler Cowen

Detail of "The Architect," Pejac (2020).

Wastebook

“…a toxic cocktail of Nextdoor, InfoWars, Fox News, Joe Rogan and thinking that empathy and reading entire books is gay.” Ben Collins
“I’ve got stickers on my laptop older than you.”
“People don’t like wearing things on their faces and don’t trust those who do” Neal Stephenson

Logoff

Install Claude Code. Ask it to do your next task. Do that for a week. Report back.

A note to email subscribers: I’m moving the newsletter from substack to micro.blog. It might be a little rough and weird. I did some formatting, but it’s not done yet. For example, the content is a little wide, eh? It’ll iron out. We’ll see. Eventually it’ll work. Probably.

The team working on the newsletter migration.

Kubernetes is the bottleneck, and you can get through it with a platform

Thu, 26 Mar 2026 11:04:15 +0200

Photo by Daniel Bryant.

Another great KubeCon talk from Abby Bangser imploring the Kubernetes people to remember how important platforms are: they make all of his stuff usable for developers and valuable to the organizations that run it all. Without a platform, you’re just put another bottleneck in place.

Also, see the go she started, along with others, at codifying what makes a good platform, at a technical level, here.

Kubernetes is the bottleneck.

She would never really put it that way, and maybe even disagree. Abby is the positive, helpful and hopeful voice of platform advocacy. And better: accurate and true.

Some of us who slogged through and lost the “container wars” of the 2010’s can’t shake off defeat and the reversion to CaaS-chaos. Which is to say:

“Kubernetes is the bottleneck” is my way of saying it.

Here’s more of my rewording in through the jaundiced eyes of the salty. Whatever type of infrastructure you have (in the modern day, IaaS and CaaS) is going to become a bottleneck unless you put a good platform in place. It’s going to harm more than help if what you want is to use software to run your organization. Once your developers 10x+ their speed of releases with AI, this bottleneck is going to become glowing red-hot obvious.

Infrastructure as bottleneck happens over and over. If you’re building out and/or running Kubernetes, pay attention to what she’s telling you - and see part one of Abby’s ongoing platform engineering thinking in her Atlanta talk, another important talk to watch if you haven’t.

Get a platform now. Layer it on-top of Kubernetes if you feel like you need Kubernetes. Having a platform on-top will save you years of puttering around at the CaaS level, millions of dollars, and that pushing on a bruise feeling of IT transformation failure.

You’re going to end up with a platform no matter what. It can be an accidental platform that just becomes another hunk of tech debt and “we spend 80% of time on maintenance” - more likely, dozens to hundreds of little accidental platforms that embody the yoke of “not my fault, now my problem” for ops…or it can be a purposefully built one.

VMware/Broadcom at KubeCon EU 2026

Thu, 26 Mar 2026 10:27:01 +0200

Here’s Claude’s take on VMware’s stuff at KubeCon - just some light editing for me.

KubeCon + CloudNativeCon Europe 2026 ran March 23-26 in Amsterdam. Here’s what VMware by Broadcom announced.

VKS 3.6 Ships

The VKS stack as seen at VMUG Connect Amsterdam 2026.

VMware vSphere Kubernetes Service 3.6 shipped with Kubernetes 1.35 support, RHEL 9 compatibility, declarative performance tuning, and improved upgrade safety targeting enterprise platform teams. The day-two operations framing is the key story - VKS 3.6 targets upgrade bottlenecks, lifecycle management headaches, and performance tuning that become problems as organizations actually scale Kubernetes, not just deploy it.

Specific operational improvements:

Declarative TuneD profiles for safe kernel and sysctl tuning for databases and high-throughput workloads, without requiring host customization
AppArmor profiles managed as Custom Resources, automatically synced across worker nodes or specific node pools
Workload cluster owners can now generate support bundles without vCenter credentials, reducing friction between Kubernetes and infrastructure teams
nftables backend support for kube-proxy on Linux nodes, improving performance and scalability
Centrally managed node-level firewall rules across all supported operating systems via API

_[If you didn’t get the memo last year, Tanzu no longer does the Kubernetes distro for VMware. That has been moved to the core VMware division at Broadcom (often called “VCF”). Tanzu now just does the Cloud Foundry based PaaS, Tanzu Platform, Spring for Java, and a whole host of databases and services. If you want Kubernetes, talk with the VMware people. -Coté]

CNI Ecosystem Opens Up

VKS now supports a path for Container Network Interface partner plugins, with Cilium and Calico currently in the validation phase.

New ecosystem partnerships with F5, Kong, and Tigera expand platform networking and security capabilities.

Velero Donated to CNCF Sandbox

Broadcom submitted the Velero backup project to the CNCF Sandbox, with the application filed in February, kicking off a process to bring the project under vendor-neutral governance with a wider contributor base. CNCF CTO Chris Aniszczyk noted that as organizations scale cloud native workloads, the focus is shifting from simple orchestration to long-term resilience and data management.

NGINX Replacement Push

Broadcom is positioning to help organizations move away from NGINX, which is deprecated in the latest Kubernetes release. No specific replacement was named in the coverage, but it’s an opening they’re clearly moving into.

Sources

The end of the meat-mouse

Tue, 24 Mar 2026 12:34:20 +0200

The agency in agentic AI feels a lot more like giving the users - the humans - agency they didn’t have. That’s what’s making it useful for me, from sorting out dumb-shit home-networking incompatibilities, figuring out taxes, and otherwise sorting my shit out. When you unleash something like Claude code on all the messy and neglected rooms in your life, you start to clean-up and pay attention more.

There’s a very bottoms-up thing here. Individuals with these tools can get a lot done. Using the tools where they are not is like assembling and continually using a Rube Goldberg device. It’d be impossible to scale to a large organization, let alone sort out all the data access and security concerns.

So far, it’s the individuals who get the productivity benefits of agentic AI. For “home use,” this stuff is killer.

I can image CEOs coming back from a week’s vacation where they downloaded Claude code or started using Claude Cowork, disappeared into all that agency, coming back to work, and telling everyone to cancel their meetings to figure out how to give everyone in the organization that kind of agency.

That happened with iPhone around 2010 or so. They were strictly a consumer device, then executives used them, and suddenly, there we were with them as the primary business device.

I don’t like the cliche line that “technology is easy, culture is hard,” but I can see where there’s something of that with enterprise AI ROI.

First, there’s a mindset and way of working that depends on a lot of human work and review.

Second, there’s probably a lot of people who still would balk at reading AI output (in docs or slides).

Third, there’s a lot of “collaboration” (meetings, reviewing slides, “email me your deck”‘ing) going on that is more about getting people informed and “synced up” than getting work done.

A person with a full-blown agentic rig doesn’t care about any of that. They have a task, they ask the agentic rig to get it done, guides it along, and then the thing is done.

We’ve built a computer out of meat-sacks called “the enterprise.” Executives use these “knowledge workers” as meat-mouses and meat-keyboards. Each of those people can be a mini-CEO now.

Or, put another way, all that prep-work and bubbling up can be done in minutes. The big meeting can be able understanding the problems, ideas, testing assumptions and fixes, and actually doing work beyond building slides, pulling CSVs from the creek CRM, visiting with customers more…doing work that doesn’t require the Office toolchain.

Sure, there’s “training” that needs to happen. You have to train people to think iteratively like a programmer, big picture like a management consultant, project management and keep the agent focused like a good manager/executive.

There’s of course securing and controlling this technology (like any other technology). I think the need to clean and prep data is likely way over blown: it just works with what you have, especially if it’s in well understood, plain old just work things like SQL, CSV files, or whatever. In fact, I’d wager the older the data format/service, the better it works.

I think a lot of the failure to find ROI in AI is because very few enterprises have actually started. Just putting a chat app in place isn’t much at all. And all those studies are probably based on things that happened in 2025, at best.

The only thing enterprises can do now is just start to try and experiment. Until they figure out how to fit AI into how their organizations works - into their “culture” - they’ll just get low-level benefits. As my pal John Willis put it, they need to define what agentic AI means for them.

That seems like a swirling mass of chaos and naive enthusiasm, sure. But, it also sounds like a better option than us all sitting in meetings, building decks, and being a meat-mouse for executives.

Here’s your assignment, which I took from Brandon Whichard. Install Claude code on your machine. The next time you need to do anything, ask it to do it for you. It won’t cook eggs for you, but most anything a “knowledge worker” would do, it’ll do. Do that for a week. I think you’ll find more human agency than you thought was possible.

The Mythical Agent-Month, Retirement Cartoons, and Coffee Brains - Related to your interests, Monday

Tue, 24 Mar 2026 08:10:21 +0200

Spotted in the Tenerife South Airport, February, 2026.

Related to your interests

Broadcom Empowers Platform Engineers to Accelerate AI and Modern Application Innovation on Kubernetes - Round-up of VMware Kubernetes releases and announcements at KubeConEU 2026.
Strengthening the Cloud-Native Ecosystem Through Upstream Collaboration - VMware’s open source work in the kubernetes community.
Broadcom expands Kubernetes support with VKS upgrades, open-source contributions and new partnerships - analysis thereof.

The Mythical Agent-Month - “Put another way, agents are so good at attacking accidental complexity that they generate new accidental complexity that can get in the way of the essential structure that you are trying to build.” // There’s something in this: if you use your productivity gains to do more instead of going home early, you end up with the same amount of problems as always; there’s just more software.
Full-stack vibe coding made easy - Seroter shows off Google AI Studio.
Claude Code Channels Lets You Text Your AI Coder From Telegram and Discord - “The concrete version: you start Claude Code with a –channels flag. Behind the scenes, it spins up an MCP server that polls a messaging platform (Telegram or Discord at launch) for incoming messages. When something arrives, the message gets injected into the active session as a event. Claude processes it with full access to its toolkit: file reads, code edits, test execution, git operations. When it finishes, it pushes a reply back to the messaging platform using a dedicated reply tool.” // Adding this kind of thing to Cowork so the norms can do it - probably in Google Chat, Slack, and Teams will be the big ROI payoff.
Claude picks the first idea that works. Make it pick the best one.
You’re still designing for an architecture that no longer exists
Use Goose with Your AI Subscription
Trusted AI for Regulated Industries - “Infosys partnering with Anthropic to go after regulated AI in financial services, healthcare, and government. The thesis: enterprises in regulated industries won’t adopt AI from the usual suspects (OpenAI, Google) because they need audit trails, explainability, and compliance guarantees. Whether ‘trusted AI’ wins over raw speed is the bet. Anthropic’s positioning here is interesting - partnering through system integrators rather than going direct.”
What Is the PARK Stack? - PyTorch, AI models, Ray, Kubernetes. // As said, still needs data, agent orchestration, and agent memory management. Those are some, you know, pretty big gaps.
Studie: Mehrheit der DACH-Unternehmen hält Cloud-Abschaltung für realistisch - 83% of DACH companies** consider it realistic that a cloud provider could unilaterally restrict or shut off access to critical IT services - the “kill switch” scenario. But only 57% have an exit strategy**. Nearly half have no contingency plan at all. (Translated by Claude)
Australians urged to work from home, drive slower to offset oil crisis shock from Middle East conflict - “International Energy Agency recommends working from home, driving 10km/h slower, public transport to offset oil shock from Middle East conflict”
Your coffee addiction may be doing your brain a favor - “Before anyone starts mainlining espresso shots in the name of science, the apparent benefits weren’t tied to heroic levels of caffeine intake, just to steady, mid-range consumption - roughly two to three cups a day - suggesting that consistency matters more than turning yourself into a walking coffee bean.”
When Your Phone Makes Time Disappear - “Frequent phone use also changes how we experience boredom. When we are so used to the fast-paced digital environment, going screenless literally makes time feel slower. That’s why even twenty minutes of focused work feels so hard after Reading Week: your mind thinks it has been way longer. This perceived slowness is why we then reach for our phones, which feel far more interesting than our high-volume, low-reward textbooks.”
CARTOON REWIND - Classic Cartoons 24/7 - Retirement plans.

You can never get enough “training” in the annual budget, and you never put enough in.

Wastebook

“Do work, go home.” Wait it out
Nancy Wheeler looks like female Rambo.

ICYMI

New Token Machines - Software Defined Talk #564 - “NVIDIA GTC, token machines, token budgets, and an AWS outage that may or may not involve AI. Plus, Matt reviews The Wizard of Oz at The Sphere.”
It’s real, enterprises just need to do the CISO work and SRE work

Logoff

I’m around KubeCon EU this week. The weather is great so far…

Tue, 24 Mar 2026 05:41:39 +0200

Using AI to help with SRE, ops, etc.:

The problem, he said, is that Claude “will get wrong correlation versus causation.” It’s like a new joiner on the team, they will think “oh, it’s a capacity problem, when actually you lost your cache.” “This is why we can’t trust LLMs for incident response,” said Palcuie. The problem is its inability to “step back and start discerning between causation and correlation… For us humans, it is hard as well.”

And:

The Jevons Paradox, said Palcuie, is “the favorite paradox in the AI industry. It’s when technological improvements increase the efficiency of our resources used, but the resulting lower cost causes consumption to rise rather than fall.”

In the case of software, “it’s easier to write software, so we write much more of it, so the complexity goes up and not down, which means things break in more interesting ways, which means more incidents, more on call… all the improvements in the tooling will be cancelled by this ever-growing complexity.”

From: Fixing Claude with Claude: Anthropic reports on AI site reliability engineering

Tue, 24 Mar 2026 05:40:11 +0200

Using AI for security log analysis and how to fix it suggestions:

Building on that foundation, leading the list of announcements is a strategy described as an “agentic security operations center” powered by its latest Gemini AI models. Google is introducing adaptive AI agents that can investigate alerts, synthesize intelligence and assist in remediation workflows in real time, which replace the need to rely on static playbooks.

The new Triage and Investigation agent is designed to autonomously analyze alerts, gather supporting evidence and deliver reasoned verdicts to help security teams reduce response times and cut through growing volumes of false positives. The new agent is now available in preview within Google Security Operations.

Google is also extending its agentic approach by allowing customers to build their own enterprise-ready security agents through support for Model Context Protocol servers. The capability removes the need for organizations to host their own MCP client infrastructure to deliver unified governance and control over custom agents, with general availability expected in early April.

From: Google Cloud unveils agentic AI security strategy with Wiz integration and threat intelligence upgrades

Enterprise ROI: Banking edition

Tue, 24 Mar 2026 05:37:34 +0200

Maybe no new banking businesses from AI yet:

“It is a long, expensive and risk-constrained transformation,” Wells Fargo analyst Mike Mayo wrote in a note to investors last week. “Indeed, we can’t find one extremely noteworthy new product or service from AI in these very early days (it’s been a little boring from a product standpoint).”

But, when it comes to running the bank:

Jamie Dimon, the CEO of JPMorgan, the biggest U.S. bank, told Bloomberg last October the bank’s annual AI-generated savings of $2 billion is about equal to what it spends yearly on the technology.

And:

BofA has said AI-powered virtual assistant Erica handles the work of about 11,000 people, and all 18,000 of the bank’s software developers use coding agents to optimize the development process, boosting productivity by about 20%. The bank has also rolled out Salesforce’s Agentforce, which enables the creation of AI agents to handle tasks, to 1,000 financial advisers, Athanasia said. “It is a long, expensive and risk-constrained transformation,” Wells Fargo analyst Mike Mayo wrote in a note to investors last week. “Indeed, we can’t find one extremely noteworthy new product or service from AI in these very early days (it’s been a little boring from a product standpoint).”

But, when it comes to how you run the bank:

Jamie Dimon, the CEO of JPMorgan, the biggest U.S. bank, told Bloomberg last October the bank’s annual AI-generated savings of $2 billion is about equal to what it spends yearly on the technology.

And:

BofA has said AI-powered virtual assistant Erica handles the work of about 11,000 people, and all 18,000 of the bank’s software developers use coding agents to optimize the development process, boosting productivity by about 20%. The bank has also rolled out Salesforce’s Agentforce, which enables the creation of AI agents to handle tasks, to 1,000 financial advisers, Athanasia said.

From: AI in banking ‘not a silver bullet’: analyst

Garbage Chairs of Amsterdam

Sat, 21 Mar 2026 09:11:20 +0200

Hidden, but seen.

Opinionated Platforms, Private Models, & Essential Dev Tooling

Fri, 20 Mar 2026 20:46:51 +0200

Related to your interests

Why Your DIY Kubernetes Stack Won’t Survive the Era of Agentic AI - 🤖 DIY Kubernetes stacks built for stateless web apps can’t handle the GPU scheduling, model versioning, inference routing, and multi-tenancy requirements of agentic AI workloads.
QCon London 2026: Morgan Stanley Rethinks Its API Program for the MCP Era - “She then demonstrated deployment gates: build-time validation that catches incomplete architectures before anything reaches production. Missing image versions and missing security controls on relationships: the CALM CLI flags all of it using structural validation with Spectral rulesets. Niculcea was direct about the impact: Morgan Stanley’s first API took roughly two years to reach production. With CALM and automated security approvals, that’s down to one or two weeks.” // And, good “turns out opinionated platforms are good” vibes: “Asked by an audience member how this approach has changed the culture among developers and architects, Gough openly acknowledged the trade-off. Developers do lose some flexibility when the platform encodes opinions about security and deployment. But they gain a working production baseline from day one that already passes every gate. The patterns can also bootstrap entire projects, pulling in frameworks like Spring Initializr or Quarkus Start, preconfigured with the right defaults, so teams can start from something already running in production rather than a blank slate.”
The Private AI Model Explosion - “In five years, 70% of the revenue created by AI will be in private models, not public models. “
How AI agents will reshape digital workplace IT operations - “Nearly 1/3 of organizations will achieve autonomous operations for 80% of their digital workplace services by 2030, up from 0% in 2025,” Gartner // I’m not sure enterprise IT can work that fast to get it setup, reliable (perf. wise), and secure, but we’ll see.
What the Best AI Users Do Differently - and How to Level Up All of Your Employees - As with all new technologies, you have to adapt how you work to it to get full benefit.
Interview: Huy Dao, director of data and machine learning platform, Booking.com - 🤖 Booking.com ML platform director on real enterprise AI practice. Practitioner talking, not a vendor. Covers data platform architecture, ML at scale for travel/hospitality.
Thoughts on OpenAI acquiring Astral and uv/ruff/ty - 🤖 When your essential dev tooling is owned by a company whose core business is something else entirely, you’re dependent on their continued goodwill.
Blending Chat with Rich UIs with Spring AI and MCP Apps
Fragments: March 19 - This is forgotten dream: “By watching how software is used, we can learn about what users really want to achieve, these observed requirements are often things that never popped up in interviews and focus groups.”
LLMs predict my coffee
StopTheMadness Pro and StopTheScript Extensions for Safari - Yes, this extension is fun. It’s like a big toolbox full of tools you hardly ever use, but that one time you need that one weird tool, you’re glad you kept it for years.

Wastebook

All jobs are software,” Brian on Reasoning.
“Washing clothes is truly the worst chore without modern technology, bar none” We can all agree.

ICYMI

New Token Machines - Software Defined Talk #564 - This week, we discuss NVIDIA GTC, token machines, token budgets, and an AWS outage that may or may not involve AI. Plus, Matt reviews The Wizard of Oz at The Sphere.
It’s real, enterprises just need to do the CISO work and SRE work

Logoff

See you next time!

It's real, enterprises just need to do the CISO work and SRE work

Fri, 20 Mar 2026 10:14:31 +0200

After using Claude more and more for task in my personal life, my current zinger analyst take on the Squawk Box would be: “OpenAI talks about business strategy, Anthropic just does it.” It’s really getting close to a sci-fi personal assistant. It takes A LOT of work to get your rig (or “harness”) setup, and to continually tune it, but it’s amazing.

Once CEO’s get their hands on this for a week, and the IT departments and CISOs figure it out, it’ll be amazing in enterprise life…again, not just for programmers, but for everyone.

The biggest benefit will be less meetings, less having to be a meat-mouse for executives who can’t or won’t make their own slides, and finally getting CRM and ERP “business analytics” nailed.

You know, until we get used to productivity benefits in 6 months and then just hate it like we do all enterprise software, because we want the next impossible thing.

It was a rather stressful period

Thu, 19 Mar 2026 21:06:10 +0200

Related to your interests

Your database is about to become an AI tool. Is it ready?
Banks struggle to scale AI as legacy tech devours IT budgets - “Executives said only 29% of annual IT budgets are set aside for transformative technologies while a larger chunk - 43% - is devoted to maintaining legacy systems.”
The AI-Washing of Job Cuts Is Corrosive and Confusing
Hardened containers for AI - “Lorenc was blunt: The bottleneck in modern software isn’t generating code anymore; it’s trust. In his keynote, he describes how AI is collapsing exploit development timelines from months to hours. AI’s sheer speed makes it impossible for defenders to rely on manual patch cycles.”
Oracle Uses JavaOne to Position Java for AI-Era Development - “For enterprise developers, that is probably the right pitch. Most Java shops are not looking for a grand manifesto. They want to know whether the platform can keep improving performance, modernizing the language, and fitting into AI-heavy architectures without forcing a rewrite or a cultural conversion.”
Inside DOGE’s early days of pressure campaigns, rule breaking and ‘chaos’ - “It was a rather stressful period and we were all under a lot of pressure.”

Wastebook

“regex surgery on markdown”
“No. We’re saving a child by kidnapping him.”

Logoff

See you next time!

Wed, 18 Mar 2026 23:43:41 +0200

Internet Coin-op