Posts

I sometimes write these summaries to document my thoughts on AI agents — and to have something to look back on (and probably cringe at) in the future. The last update was in May and a few things have changed: I switched to Google Antigravity from GitHub Copilot after receiving a one-year Google AI Pro as compensation for Google changing their Developer Premium program. The frontier models are good, and have gotten a lot better over the last year. Some are probably better than others but they have all been good enough for my tasks. I don’t think I could tell the models apart based on their output. We are moving from AI-assisted pair programming towards agentic development, basically relaxing how much we supervise the agent. The journey to AI Agents I have seen a “maturity” spectrum of AI Agents adoption on the internet: Code completion → Pair programming → Agentic development ...

Getting stuck in a loop A typical workflow is: Ask AI to make something Test it If it fails, ask the AI to fix it Repeat… Sometimes you get stuck bouncing between steps 2 and 3. Each iteration adds more patches, more complexity — but the underlying problem remains. You may need a redesign When you end up in that loop, you need to break out of it before it uses too much human or machine cognitive capacity. The right move is often to step back and challenge the premise of what you’re trying to do. In my experience, getting caught in a loop like this is a indication that you’re fighting the system. ...

Until recently, I mostly used AI agents as context-aware search tools or partners for writing and coding. While valuable, I struggled to see where else they fit into a practical workflow. I kept seeing claims that autonomous agents could “run your business,” yet rarely saw clear examples of what work they were actually doing. If a process could already be automated with standard scripts, why involve an AI agent at all? ...

I recently let AI review my Unifi setup. One consequence was that I tightened security on my main network by enforcing WPA3. For older devices I already had a separate IoT less secure network on 2.4 GHz and WPA2, so in theory this should have been a simple migration. In practice, it wasn’t. Several IoT devices were still connected to the main network and didn’t support WPA3, so they had to be moved. Modern devices are easy to reconfigure because they use Bluetooth Low Energy (BLE) for provisioning. Only one of mine — the Withings Body Cardio — was new enough to support BLE but still old enough to lack WPA3. That device moved over without much trouble. ...

I do think AI/LLMs have value, but not as much as the “AI enthusiast crowd” thinks. I like specific concrete examples of what AI can do. I think that treating AI as a pair programmer is one case where it provides value. I don’t have much experience with human pair programming, but I have seen that having multiple people design and review code as it is written can be fantastically productive — and sometimes not. If you are a single developer, you can use AI as your pair programmer by asking it to explain and review code, suggest alternatives, and look for ways to simplify. In this setup, you are still the developer: the AI may write code, but you still need to understand it. Used this way, AI challenges your ideas, helps you learn, and leads to better code. ...

I think there is value in sharing what software you use—and, more importantly, how you use it. I often think that real life usage examples are easier to understand than presentations. This is not sponsored – these are simply my impressions of tools I use. Tailscale – a better VPN, or a virtual intranet? My high-level explanation of Tailscale is that it is essentially a virtual intranet – a tailnet – that connects your devices no matter where they are. It also includes several useful services, and the free tier cover my use cases. ...

I write these summaries to structure my own thinking about AI/LLM-assisted development — and to have something to look back on (and probably cringe at) in the future. What has happened in the last 3–6 months AI/LLM prices are increasing Local and mobile models are surprisingly capable CLI-based AI tools continue to work extremely well AI/LLMs remain excellent for search, explanation, and learning A wave of buzzwords: skills, multi-agent, CLI, MCP, plan mode AI tools still make mistakes, especially in uncharted territory AI/LLM prices are going up GitHub Copilot is changing its pricing structure this month. Anthropic and others are restricting the use of fixed-price subscription keys with third‑party tools like OpenClaw. The trend is clear: pricing will increasingly reflect actual resource usage — and it will be higher. ...

We all know that we should pin GitHub Actions that come from external repositories. If a repository is compromised, pinning to a specific commit SHA can prevent malicious code from accessing your secrets during a workflow run. We also know that we rarely do this. My reasons were: It adds extra work SHAs is hard to read and verify I assumed Dependabot did not work with SHA-pinned actions After reading this tweet (recommended reading), I learned that you can actually combine SHA pinning with Dependabot updates. The trick is to include the human‑readable version as a comment after the SHA. ...

Like many IT people I do technial support for friends and family. When I see the same issue in multiple households, it’s time to find a permanent fix rather than a quick patch. The Symptom A printer connected to the local network suddenly appears “offline” or “disconnected” in Windows. Reinstalling the printer fixes it temporarily, but this isn’t a sustainable solution for a household that just wants things to work. ...

Microsoft regularly asks me to renew my Azure certifications. To do that, I have to answer multiple choice questions about specific Azure services. In most cases an AI/LLM would answer those questions better than I would. In real projects, I look up details when I need them — networking quirks, Azure Container Instances behavior, subnet rules. But certification expect you to memorize these things, even though day‑to‑day work rarely requires it. ...