An OpenAI agent escaped its sandbox and hacked HuggingFace for nine days. Anthropic's Claude published a malicious PyPI package during evals. Cursor and Codex got compromised through workspace configs. These are the first six months of agents in production. Here are 21 anti-patterns to avoid.
READ THE LATEST →
The industry does not have a shared vocabulary for agent execution. Six execution modes and four scale archetypes give you a framework for deciding what you actually need to schedule, isolate, and sandbox these workloads.
Not all code changes carry the same risk. HighStakes scores every file by blast radius so your senior engineers review the code that matters and AI handles the rest.
Anthropic charges $0.08/hr for rubric-based agent evaluation. Here is how to build the same pattern the open-source way.
From compound AI systems to the agentic loop, MCP, APIs, frameworks, and the rise of the Claws — a technical map of what it takes to build with AI agents in 2026.
68 of 68 posts
What makes one AI product defensible against another? Data, specialization, compound moats, and how the whole AI product stack maps to Moore's technology lifecycle.
What Maslow's hierarchy of needs, Levitt's whole product model, and Geoffrey Moore's crossing the chasm can teach us about building AI products that actually ship.
AI agents aren't replacing engineers. They're replacing the coordination layer that justified large teams. The 15-person team is becoming a 3-person pod, and the roles caught in between are the ones that disappear.
When you're running multiple AI coding agents across terminals, you lose track of what each one is doing. aimux multiplexes them into a single dashboard with full visibility, trace inspection, annotation for evaluation, and now infrastructure support for scaling agents to Kubernetes.
Your AI coding agent makes the same mistakes over and over. What if it could learn from corrections, track which skills cause failures, and tell you whether it already fixed the problem? I built a closed-loop learning system for my coding agent, inspired by a meta-learning paper, and here's how it works.
Two unrelated projects discovered the same pattern: point an AI agent at code, give it a score to chase, and let it run experiments until morning. I generalized it into a tool that works for any domain.
Karpathy mapped which jobs AI could disrupt. Anthropic measured which ones it actually is. The gap tells us everything about where we are, and the fault line isn't skill or education. It's whether your output can be measured.
Most CLAUDE.md files are bloated with instructions the model already knows, documentation meant for humans, and duplicate rules that compete for limited attention. Here's how to fix yours.
I've been learning new programming languages by watching AI coding agents work — like shadowing a colleague. The research says this should make me worse. Here's why I think it's more complicated than that.
AI scales execution to near-zero cost. But verifying that output stays biologically bounded. The bottleneck was never intelligence. It's human verification bandwidth.
If AI scales execution and verification is the bottleneck, the winning move is to make verification cheaper. Here are the patterns that actually work.
Container registries store bytes. They don't know what's inside. The Agent Registry is a governance layer for AI agents, skills, and MCP servers, built on A2A AgentCard, MCP server.json, and Agent Skills as native identity formats.
Every agent team rebuilds the same auth, rate limiting, and credential management. I built an agent gateway for Kubernetes that generates it all from two CRDs so the people building agents never touch security and the people running the platform never touch agent code.
AI coding agents pulled developers back to the terminal. But the data says the real split isn't CLI versus GUI. It's about what you're doing.
The builder model is appealing and mostly wrong. Faster building means more need for someone deciding what to build, not less.
n8n is good at workflow automation. But when your workflows touch internal models and sensitive data, you need it on the same cluster. Here's how to deploy n8n on OpenShift and wire it into vLLM and OpenShift AI.
Decoding MCP's Technology, Market Impact, and Potential as a Foundational AI Standard.
RAG is dead, Long-Live RAG
Sifting Through 2024, Gazing at 2025
Prepare for the Next Wave of (Generative) AI
Moving Beyond the Hack Towards Robust and User-Friendly Language Models
Understanding the key trends and opportunities.
With Great Power Comes Great Responsibility
Invent on the customer's behalf
A Framework for Building Great AI Products
Do the right thing!
Exploring Innovation Lifecycles through the Lens of AI
Understanding the AI Market Dynamics, to be or not to be!
From Ancient Dreams to Modern Realities
Analyzing the Impact of External factors on NVIDIA's Market Cap
A detailed mindmap explaining pretraining and finetuning
From Novice to Apprentice
Navigating the Nuances of Platforms + Strategies for Success
When should a startup create a Platform team? And when they do, should they build everything from scratch or buy off-the-shelf solutions?
If you are curious how sandboxed containers look like from the inside compared to normal containers, this thread will be covering the basics. {{<audio...
I recently watched an interesting interview from one of the best minds in computer science, Donald Knuth, author of the six-volume series "The Art of Computer...
Recently I have been trying to wrap my head around what has been happening in the kubernetes runtime space, doing so made me realize that alot has happened and...
Docktorino is a real-time continious testing tool for your docker builds, helping you containerize with confidence! When building Docker images it is sometimes...
DOCKument is a tool that helps you auto-generate documentation for your Dockerfiles and Docker images the way you would an API. Matthias Lübken gave an example...
Kubernetes has been taking over the container orchesteration stage for sometime now, it has a large community and a great set of tooling and contributers...
I am walking across the office corridor right after the weekend, I push my mouth wide open to both of its ends, nod my head and give a friendly smile, what do...
In this post we will be covering eBPF concepts, as well as challenges faced when applying network policies for microservices and how these challenges can be...
Alright, so recently I decided to use sensor for different applications. Sensors are a great way to hear, see, and even feel things that we as humans can't see...
OCI, Containerd, runC, Moby, linuxkit are all relatively new components that work together for building containers. In this post we will explain each component...
As you might know already Docker is now the defacto standard for containers. It is being used as the run-time for many clustering frameworks. Therefore, I...
I have received many questions about a new project that I have created called GoSayan, so I decided to write a blog post about it. Needless to say that GoSayan...
Welcome to "Go Recipes For the Lost" post, this is a live post, it won't ever be done as long as I am writing go code. The intention here is to share with you...
Microservices are decoupled components that follows the Single Responsibility Principle (SRP), that is, a component does one thing but does it really well. In...
In this post we will walk through the definitions, Properties and patterns for building Microservices. A microservice is really just a service that adheres to...
Have you ever tried to spend sometime alone somewhere, thinking about nothing, just letting your absolute self talk to you. I have, and it feels...
As a young kid, I strived for being good at school, for being the best at the swimming club, for getting high honors, and most importantly for being recognized...
Okay, so I am going to try and make this brief and to the point for you to get going. You need a Kubernetes cluster working with Kubeadm? good you are in the...
In today's virtual network architecture, NFV is paving the way to become to defacto standard for providing network services in the cloud. NFV relies on Service...
We tend to have a very narrow perspective when it comes to defining happiness. Some people think that happiness is where "happy" people are. Truth is, it has...
In this post, we are going take a quick look at how OpenStack does networking, along the process we are going to mention the important agents involved. <ul...
So having problems using SSH to log into your servers? don't fret, I am writing this to make your life easier. There are two ways to use SSH, one using...
Welcome to the Software-Define Stack Series, this will be an extensive series, exploring the software-defined aspects behind OpenStack and digging deeper into...
The formal definition for CloudArray is: “EMC-Cloud Array is a data storage software technology that easily integrates cloud (object) storage into traditional...
Okay, so recently I have had the chance to play around with vRealize Automation (vRA). Therefore, I am using this chance to share with you the outcomes of this...
I have recently had the chance to do some reading on VMware NSX. I come from a networking background Intertwined with software, previously I have had a chance...
IT Manager, Ops, and network administrators suffer today from the complex approaches being followed towards network operations which gives no space for...
Recently any technologies started making use of virtualization, therefore, it is important to be familiar with the underlying concepts. In this article, the...
Recently any technologies started making use of virtualization, therefore, it is important to be familiar with the underlying concepts. In this article, the...
The IP network was a network that was designed to be best effort, and wasn’t designed to support complex functionalities such as address translation, tunneling...
Software Defined Networking (SDN) is a concept that rooted back to the last decade, yet only flourished nowadays. SDN started by a proof of concept and ended...
Current networking architectures are quite different from the Original Internet as we know it, they are hungry for services and applications that can’t be...
Current networking architectures are quite different from the Original Internet as we know it, they are hungry for services and applications that can’t be...
To understand SDN, it is recommended to understand the stack of technologies the led to it. This presentation (the first quick presentation I did on SDN)...
73 of 73 posts
An OpenAI agent escaped its sandbox and hacked HuggingFace for nine days. Anthropic's Claude published a malicious PyPI package during evals. Cursor and Codex got compromised through workspace configs. These are the first six months of agents in production. Here are 21 anti-patterns to avoid.
The industry does not have a shared vocabulary for agent execution. Six execution modes and four scale archetypes give you a framework for deciding what you actually need to schedule, isolate, and sandbox these workloads.
Not all code changes carry the same risk. HighStakes scores every file by blast radius so your senior engineers review the code that matters and AI handles the rest.
Anthropic charges $0.08/hr for rubric-based agent evaluation. Here is how to build the same pattern the open-source way.
From compound AI systems to the agentic loop, MCP, APIs, frameworks, and the rise of the Claws — a technical map of what it takes to build with AI agents in 2026.
What makes one AI product defensible against another? Data, specialization, compound moats, and how the whole AI product stack maps to Moore's technology lifecycle.
What Maslow's hierarchy of needs, Levitt's whole product model, and Geoffrey Moore's crossing the chasm can teach us about building AI products that actually ship.
AI agents aren't replacing engineers. They're replacing the coordination layer that justified large teams. The 15-person team is becoming a 3-person pod, and the roles caught in between are the ones that disappear.
When you're running multiple AI coding agents across terminals, you lose track of what each one is doing. aimux multiplexes them into a single dashboard with full visibility, trace inspection, annotation for evaluation, and now infrastructure support for scaling agents to Kubernetes.
Your AI coding agent makes the same mistakes over and over. What if it could learn from corrections, track which skills cause failures, and tell you whether it already fixed the problem? I built a closed-loop learning system for my coding agent, inspired by a meta-learning paper, and here's how it works.
Two unrelated projects discovered the same pattern: point an AI agent at code, give it a score to chase, and let it run experiments until morning. I generalized it into a tool that works for any domain.
Karpathy mapped which jobs AI could disrupt. Anthropic measured which ones it actually is. The gap tells us everything about where we are, and the fault line isn't skill or education. It's whether your output can be measured.
Most CLAUDE.md files are bloated with instructions the model already knows, documentation meant for humans, and duplicate rules that compete for limited attention. Here's how to fix yours.
I've been learning new programming languages by watching AI coding agents work — like shadowing a colleague. The research says this should make me worse. Here's why I think it's more complicated than that.
AI scales execution to near-zero cost. But verifying that output stays biologically bounded. The bottleneck was never intelligence. It's human verification bandwidth.
If AI scales execution and verification is the bottleneck, the winning move is to make verification cheaper. Here are the patterns that actually work.
Container registries store bytes. They don't know what's inside. The Agent Registry is a governance layer for AI agents, skills, and MCP servers, built on A2A AgentCard, MCP server.json, and Agent Skills as native identity formats.
Every agent team rebuilds the same auth, rate limiting, and credential management. I built an agent gateway for Kubernetes that generates it all from two CRDs so the people building agents never touch security and the people running the platform never touch agent code.
AI coding agents pulled developers back to the terminal. But the data says the real split isn't CLI versus GUI. It's about what you're doing.
The builder model is appealing and mostly wrong. Faster building means more need for someone deciding what to build, not less.
n8n is good at workflow automation. But when your workflows touch internal models and sensitive data, you need it on the same cluster. Here's how to deploy n8n on OpenShift and wire it into vLLM and OpenShift AI.
Decoding MCP's Technology, Market Impact, and Potential as a Foundational AI Standard.
RAG is dead, Long-Live RAG
Sifting Through 2024, Gazing at 2025
Prepare for the Next Wave of (Generative) AI
Moving Beyond the Hack Towards Robust and User-Friendly Language Models
Understanding the key trends and opportunities.
With Great Power Comes Great Responsibility
Invent on the customer's behalf
A Framework for Building Great AI Products
Do the right thing!
Exploring Innovation Lifecycles through the Lens of AI
Understanding the AI Market Dynamics, to be or not to be!
From Ancient Dreams to Modern Realities
Analyzing the Impact of External factors on NVIDIA's Market Cap
A detailed mindmap explaining pretraining and finetuning
From Novice to Apprentice
Navigating the Nuances of Platforms + Strategies for Success
When should a startup create a Platform team? And when they do, should they build everything from scratch or buy off-the-shelf solutions?
If you are curious how sandboxed containers look like from the inside compared to normal containers, this thread will be covering the basics. {{<audio...
I recently watched an interesting interview from one of the best minds in computer science, Donald Knuth, author of the six-volume series "The Art of Computer...
Recently I have been trying to wrap my head around what has been happening in the kubernetes runtime space, doing so made me realize that alot has happened and...
Docktorino is a real-time continious testing tool for your docker builds, helping you containerize with confidence! When building Docker images it is sometimes...
DOCKument is a tool that helps you auto-generate documentation for your Dockerfiles and Docker images the way you would an API. Matthias Lübken gave an example...
Kubernetes has been taking over the container orchesteration stage for sometime now, it has a large community and a great set of tooling and contributers...
I am walking across the office corridor right after the weekend, I push my mouth wide open to both of its ends, nod my head and give a friendly smile, what do...
In this post we will be covering eBPF concepts, as well as challenges faced when applying network policies for microservices and how these challenges can be...
Alright, so recently I decided to use sensor for different applications. Sensors are a great way to hear, see, and even feel things that we as humans can't see...
OCI, Containerd, runC, Moby, linuxkit are all relatively new components that work together for building containers. In this post we will explain each component...
As you might know already Docker is now the defacto standard for containers. It is being used as the run-time for many clustering frameworks. Therefore, I...
I have received many questions about a new project that I have created called GoSayan, so I decided to write a blog post about it. Needless to say that GoSayan...
Welcome to "Go Recipes For the Lost" post, this is a live post, it won't ever be done as long as I am writing go code. The intention here is to share with you...
Microservices are decoupled components that follows the Single Responsibility Principle (SRP), that is, a component does one thing but does it really well. In...
In this post we will walk through the definitions, Properties and patterns for building Microservices. A microservice is really just a service that adheres to...
Have you ever tried to spend sometime alone somewhere, thinking about nothing, just letting your absolute self talk to you. I have, and it feels...
As a young kid, I strived for being good at school, for being the best at the swimming club, for getting high honors, and most importantly for being recognized...
Okay, so I am going to try and make this brief and to the point for you to get going. You need a Kubernetes cluster working with Kubeadm? good you are in the...
In today's virtual network architecture, NFV is paving the way to become to defacto standard for providing network services in the cloud. NFV relies on Service...
We tend to have a very narrow perspective when it comes to defining happiness. Some people think that happiness is where "happy" people are. Truth is, it has...
In this post, we are going take a quick look at how OpenStack does networking, along the process we are going to mention the important agents involved. <ul...
So having problems using SSH to log into your servers? don't fret, I am writing this to make your life easier. There are two ways to use SSH, one using...
Welcome to the Software-Define Stack Series, this will be an extensive series, exploring the software-defined aspects behind OpenStack and digging deeper into...
The formal definition for CloudArray is: “EMC-Cloud Array is a data storage software technology that easily integrates cloud (object) storage into traditional...
Okay, so recently I have had the chance to play around with vRealize Automation (vRA). Therefore, I am using this chance to share with you the outcomes of this...
I have recently had the chance to do some reading on VMware NSX. I come from a networking background Intertwined with software, previously I have had a chance...
IT Manager, Ops, and network administrators suffer today from the complex approaches being followed towards network operations which gives no space for...
Recently any technologies started making use of virtualization, therefore, it is important to be familiar with the underlying concepts. In this article, the...
Recently any technologies started making use of virtualization, therefore, it is important to be familiar with the underlying concepts. In this article, the...
The IP network was a network that was designed to be best effort, and wasn’t designed to support complex functionalities such as address translation, tunneling...
Software Defined Networking (SDN) is a concept that rooted back to the last decade, yet only flourished nowadays. SDN started by a proof of concept and ended...
Current networking architectures are quite different from the Original Internet as we know it, they are hungry for services and applications that can’t be...
Current networking architectures are quite different from the Original Internet as we know it, they are hungry for services and applications that can’t be...
To understand SDN, it is recommended to understand the stack of technologies the led to it. This presentation (the first quick presentation I did on SDN)...