Martin Matusiak

The better you know the basics the more advanced you are. Flashy tools only work if you actually know what you are doing with them. Handing a 16 year old a Ferrari race car is probably a bad idea. Handing most software dev teams an AI agent is pretty much the same thing. You are just helping them to crash faster and more spectacularly.

9

3 Comments

Of all the low cost bare metal or cloud offerings, Hetzner seems to offer the best value for money, decent customer support and good ancillary offerings. Their server auctions give you bare metals at throw away price (if you don’t want great reliability - mostly for dev work, experiments or as a playground) and their cloud offers pretty decent stability (maybe not as good as AWS/Azure/GCP but at par with many other private cloud providers). I’ve tried OVHcloud in the past, and Hetzner beats them in most areas.

23

3 Comments

Having a boring morning at work? Why not: - create a component or a service with an LLM - ask the model to review the code **like it's never seen it before** - watch it abuse its own creation mercilessly 🍿 Seriously though, this is a great way to provide some QA for generated code. PS. Also works with human devs if you extend the gap between code creation and review.

5

Staring at the screen this afternoon, waiting for my gitlab pipelines to turn green. My brain can't deal with the pause and it jumped back to a video I saw a few weeks back on General Agent's Ace. The video has been bugging me for a few weeks. Not necessarily a full-blown existential crisis, but it did raise the question; are we building our web portals completely wrong? If the next generation of AI tools can scrape screens and perform complex tasks across multiple products, websites, tools without a sweat, what does that mean for how we build them? The thought is very very tempting: could we not just sacrifice UX for raw speed? Do we even need quality interfaces? If AIs like ChatGpt and Gemini can understand our gibberish prompts, surely newer AIs can handle a clunky UI. But the idea falls apart the second the AI fails and a real, stressed-out human has to jump in to save the day. Preferably without being forced onto some hostile, machine-first interface. This leads to the real question; UI vs API. Or is that the wrong way to think? Instead of building separate machine-only APIs, what if we focused on making our UIs pleasant enough for humans, but boringly predictable for machines? Essentially, treating the UI as the API? Thinking like this would change a lot. Instead of the popular, but very “dry” OpenAPI specs; Could well-defined UX patterns become the next specification language, understood by humans and AIs alike? Perhaps. But it points to a more immediate reality: the concept of accessibility is getting a massive redefinition. For years, we've seen accessibility as a way to support human users with different needs. Now, AI agents are becoming a new class of "user" that relies on the exact same principles. A clean, semantic UI that’s clear to a screen reader is also perfectly clear to an AI agent. The business case for accessibility is suddenly no longer just about inclusion; it's about enabling automation. This feels like a shift. I'm curious to hear other perspectives. Are we about to see a renaissance for accessibility, driven by the needs of AI? What does "designing for everyone" mean when "everyone" now includes machines?

31

15 Comments

I was suggested a post on here trying to define the differences between DevOps Engineers, SRE, Cloud Engineers and Platform Engineers for people new to IT. It was pretty bad to be honest, overly simplistic and reductionist without any of the context around the positions, basically boiling them down to Pipelines, Monitoring, Infrastructure and reusable stuff. The comment section was a complete mess but one comment stood out to me defending the post saying "That these distinctions matter for division of labour." That's sat with me since I've read it. In an industry where we've stripped these titles of their original context do any of them actually matter? #SRE #DevOps #PlatformEngineering

8

2 Comments

We recently automated our test coverage process using Claude Code agents. The key was building two specialised agents - one for writing tests, one for reviewing them - rather than using single-prompt generation. I wrote up the technical implementation, including: - How agent descriptions trigger task routing - Why tool configuration matters - The 4-step iterative process we use - Lessons learned from implementation Full breakdown here: https://lnkd.in/dgCM8XEu

11

October learnings: 1. Some tasks that used to take quite some time a few years ago are now getting done fast because the tooling evolved. When I have to parse something my usual tool was ANTLR which is great. However the scaffolding for an ANTLR is quite heavy so I found alternatives that made my work quite simpler. Recently I wrote a transpiler and after some iterations on the tooling I landed on peggy.js (https://peggyjs.org/) that is based on PEG formalism. Worked nice with JS, not so nice with TS. 2. One of the things we are missing with agents and AI is the lack of true inferences. When looking at the MCP mess (Why tools are tools and why synthetic tools?) we can see that there is a need to add some kind of ontologies over the marsh of concepts in MCP. Kind of what RDF/OWL was in the early web era. But nowadays people forget about semantics and just jump in the form bandwagon... 3. People started to observe that in fact Svelte's ecosystem is larger than React's, hence the long standing argument against Svelte is not holding anymore. Svelte, that doesn't use VirtualDOM, accommodates any other JS/TS library without any adaptations. It's 'plug&play'.

9

Executable ADRs? That directly resonated with me when watching Ben's "This Week in AI" stream. We’re basically talking about architecture fitness functions. I totally see good software engineering practices becoming more relevant than ever before. (Actually, quite the opposite of what "vibecoders" pray 😬) We have to think about fitness functions in a completely new light because AI shifts the cost-benefit scale fundamentally. It has never been easier to implement them, and when we use them as the harness for our coding agents, their value actually multiplies. 🚀 I agree with Ben 100% - exploring this is far more exciting than polishing our markdown writing skills.

10

We log in every day. OAuth works. SAML sits somewhere in the background. OIDC quietly does its job. And yet, when asked why… we often pause. Same screen. Same button. Very different questions being answered underneath. I wrote this piece after realising how often we confuse identity with access, not out of ignorance, but comfort. If “it just works” has ever felt like enough, this might be worth reading. (Link in comments) #CloudArchitecture #EngineeringLeadership #TechFundamentals #IdentityManagement #BuildingInPublic

8

1 Comment

I’m broadly positive about AI, but I think it’s worth challenging one assumption I keep seeing: that we’re waiting for intelligence to arrive before we see a real productivity boom. We’ve had plenty of non-artificial intelligence in organisations for a long time. Skilled people, good ideas, hard-won experience. The issue hasn’t been a lack of intelligence — it’s been a lack of discipline and skill in applying it. AI unquestionably makes intelligence cheaper and more accessible. That’s powerful. But lower cost doesn’t automatically translate into better outcomes if leaders still struggle to turn insight into execution. No matter how easy intelligence becomes to obtain, the bottleneck remains the same: Clear problem definition Strong operational discipline The ability to translate insight into sustained action Until those improve, AI will amplify existing habits more than it transforms results. The limiting factor isn’t intelligence. It’s the application of intelligence.

1

Confused about when to use Claude Desktop versus Claude Code? Or maybe you are still just using Claude in your browser? Most people do not know the difference. That misunderstanding will cost businesses real 💰 💵 over time. Here is the simplest way to think about it: One is built for thinking. The other is built for doing. If you are using AI for strategy, content, or decision-making, you need an environment that helps you think clearly. This is where Claude Desktop and Claude.ai shine. The desktop app and browser version are designed for: • Writing and refining ideas • Analyzing information • Building strategies • Creating polished outputs like documents, presentations, and visuals It is your workspace for clarity. But when you need AI to actually execute tasks, handle files, run processes, or move work forward automatically, you need a different approach. That is where Claude Code comes in. This is your operator. It does not just suggest what to do. It helps get it done. Now here is where it becomes important for business owners. AI becomes far more powerful when it stops acting like a tool and starts acting like a team member. That is what agents really are. Think of an agent as: AI + structure + instructions Instead of starting from scratch every time, you are giving AI a role, context, and memory so it behaves like someone who already understands your business. And you can take it further. You can have multiple specialized agents working at the same time. One focused on research One on writing blogs One on onboarding new clients All moving work forward in parallel, same way like your employees do. So why does this matter? Because most businesses are only using AI to think, when they could also be using it to operate. Forget trying to write the perfect prompt. The real shift is building systems where: • Ideas are generated • Tasks are executed • Work moves forward without constant input If you are a business owner, here is the takeaway: Use the desktop app to figure things out when you are stuck. Use the terminal version to get things done and automated. Using AI is no longer the advantage. Knowing HOW to use it is. The gap between businesses will not be who uses AI. It will be how they use it. Soon, the companies pulling ahead will be the ones running multiple agents in the background to get work automated. And if you don't know how to build agents...Don't worry! I will be sharing more on how to build simple, cost-effective agents. No technical expertise will be required.

1

I once worked for an organisation that started with people who knew very little about software architecture and programming. They wrote scripts, deployed them through screens, and relied heavily on network communication and cron jobs for scheduling. As the organisation grew, new engineers were hired mainly for their problem-solving ability, not for a specific tech stack or strong programming background. This meant we had multiple developers, each using different languages. We had scripts in Java, PHP, Python, and more — yet we still managed to work together and achieve our goals. In a way, this forced the organisation to put systems in place that supported team scaling, without even realising it. The infrastructure itself was poor, but migrating to a better-architected system later was not as difficult as expected, largely because team scaling was already in place. This is very different from how many start-ups begin. Often, they start with a very proficient developer. The infrastructure is done right early on, but because they are working solo and moving fast, little or no attention is given to team scaling. I only started appreciating this recently while taking a course on software architecture. When people talk about scaling in software architecture, most think only about vertical and horizontal scaling. What is often forgotten is that scaling has a third dimension: team or organizational scaling. Team scaling is the ability of a system and its architecture to support the growth of the engineering team without significantly reducing productivity, increasing risk, or creating excessive coordination overhead. If team scaling is not considered early, we often end up with systems that scale well technically but struggle as the organization grows. Efficiency increases at first, then plateaus, and eventually declines. This shows up as more code conflicts, larger and riskier releases, delayed deployments, communication overhead, and difficult onboarding. Fixing these issues later usually costs a lot of time and money. In conclusion, it’s important to think about team and organizational scaling from the very start, even if you are working as a solo developer. Which do you think is harder to fix later: bad infrastructure or poor team scaling?

13

RAG isn’t dead, but context engineering is the new hotness So, is RAG (Retrieval-Augmented Generation) dead now? Last May I asked that question of Douwe Kiela, CEO of Contextual AI, based on the growing hype around MCP (Model Context Protocol). Both are data retrieval mechanisms for Large Language Models, but it’s MCP that has taken all the headlines over the past year. The truth is, RAG has fallen away as a term used by developers and AI engineers. Even Kiela, who co-authored the 2020 academic paper that introduced RAG to the world, admits that a trendy new term has taken over. Stay connected for industry’s latest content – Follow Deepthi Talasila #DevSecOps #ApplicationSecurity #AgenticAI #CloudSecurity #CyberSecurity #AIinSecurity #SecureDevOps #AppSec #AIandSecurity #CloudComputing #SecurityEngineering #ZeroTrust #MLSecurity #AICompliance #SecurityAutomation #SecureCoding #linkedin #InfoSec #SecurityByDesign #AIThreatDetection #CloudNativeSecurity #ShiftLeftSecurity #SecureAI #AIinDevSecOps #SecurityOps #CyberResilience #DataSecurity #SecurityInnovation #SecurityArchitecture #TrustworthyAI #AIinCloudSecurity #NextGenSecurity https://lnkd.in/gXkFArHX

1

My role requires a high amount of communcation at low latency with many teams. We use Slack like many other companies to achieve this. Many of us use the split view/window activity monitor to stay on top of incoming messages and Slack has been rolling out a behaviour change to that popped out view recently. In the previous mechanics, you'd click on an item in the view and it would open that thread in the main view. There were minor rough edges with that, but it worked. You kept the view of incoming messages in the popped activity view and updated discussion threads in the main view. After the update, clicking an item in the activity view opens the thread in. the. activity. view. If you don't navigate to that discussion in the main window or open it in a new window, your activity window is now completely useless until you click back, back, whatever, to get to the activity view. Every one of my colleagues that has experienced this change has experiencd incredible frustation. Maybe this works if you get 5 messages a day, but it doesn't work in my envornment. I personally consider it unusable. I would love a knob to at least revert to original behaviour Slack. I have provided feedback directly already, but I'm hoping that a public post might solicit a shared response from other people. Not sure if user feedback was tested before this change was made, but I've yet to find a person that thinks it is positive. (My personal thoughts, not representing my employer with this post.)

23

3 Comments

If your job involves telling other people what to do (lead engineer, architect, line manager), agentic coding workflows feel natural. Defining work for other people isn't that different than defining work for an agent. In many regards it's easier. If your job involves doing what other people tell you what to do, agentic coding is fun, but you won't experience the same multiplication effect. It's harder to scale yourself, because you've never had to before, so you don't see all the opportunities.

16

3 Comments

Monorepos? I have recently been dropped into an in-production monorepo. For anyone new to the term, it means keeping multiple projects in a single repository. There are some big wins. Sharing design systems between front-end apps is seamless. Code reuse is simple. Dependencies stay aligned and collaboration is more open. There are headaches too. Build times get heavy. Onboarding is rarely smooth. I have yet to see a README that truly covers setup. Environment variables go missing, Docker breaks, databases and Redis need wrangling. It all adds up. I have always liked Git submodules as a middle ground, but they have never really been fit for purpose. Each project should stand alone, while the monorepo should provide clear instructions for running, deploying and developing different parts of the stack. My view is that monorepos can work brilliantly if documentation stays fresh and technical debt is managed properly. In reality, most businesses struggle with both. Sometimes smaller repos that can be retired if they do not make it to production are the safer bet.

2