Royal Schiphol Group

Claude Code doesn't listen and doesn't do what it plans. It can generate code, it can make things work but today I found out (after some vibing) the code design was not as promised. The connection protocol was supposed to be swappable, this was a deliberate design principle Claude itself proposed and planned for. Yet the connection protocol was smeared all over the place, not 'pluggable' at all. I then asked it to rewrite the messy code according to Clean Code principles, it tried but ignored code obviously not in line with CC principles because, it later explained, "wasn't biting". I asked it again with more force, but it left this function in the cli.py file (the entrypoint for the app): async def _exchange_hello(url: str, identity: Identity) -> HelloAck: """Open a sync WS, exchange hello/hello_ack, return the ack.""" async with websockets.connect(peer_ws_url(url)) as ws: await ws.send ( encode(Hello(peer_id=identity.peer_id, device_name=identity.device_name)) ) ack = decode(await ws.recv ()) if not isinstance(ack, HelloAck): raise PairError(f"expected hello_ack, got {type(ack).__name__}") return ack Every single line has issues that make this not a candidate for cli.py The cli file should deal with turning command line arguments into actions, not dealing with web-socket realities - even Codex GPT 5.5 didn't see this is a concern until I explicitly pointed to it. What's going on here? Current LLM coding tools are geared towards minimal diffs, per-line review and a strong tendency to avoid creating new files/classes. Perhaps the result of training on millions of Github issues in open source projects? My fourth attempt: I asked Codex to create more classes, but Codex went overboard and created one class too many, a PairingClient which had only one function and invisible from the outside. I have a hard time believing we can SKILL.md ourselves out this kind of behaviour with coding LLMs pre-optimized for minimal diff and both resisting and overshooting when asked to mind design principles. Another concern is my own limitation: I would not be able to verbalise every intuition I have about programming up front. #LLM #Codex #Claude

Still buzzing from two amazing days at #ECS2026 in Cologne 🇩🇪 last week! This was my very first ECS, and getting to be there as a speaker made it even more special. I’m genuinely proud that I had the chance to present a session at an event like this 🙏 What made these days so memorable was the mix of learning, inspiration, and great conversations. I got to attend some fantastic sessions, including David Lorenzo López talk on Dataverse and Business Central, The Virtual Entity Way. Really cool to see how virtual tables can be used to connect Business Central, Dataverse, and Power Pages in such a smart way 🚀 Charles Sexton session on accessibility in Power Apps was another highlight for me. It was great to see real attention being given to such an important topic. Building accessible applications should never be an afterthought 💯 And ending the event with a session from two Dutch friends, Daniel Laskewitz and Miguel Verweij (also my travel buddy, thanks for the fun trip ❤️ ), was just awesome. It was also great to see that there is still real attention for the power of AI Builder 💪 But as always, events like this are about more than just the sessions. Catching up with familiar faces, meeting new people, sharing ideas, and having great conversations is what makes these community events so valuable. Paul Beck , it was so great to see you again and to talk about all the cool stuff you’re doing with #Playwright ! Huge thanks to everyone who made these two days so much fun. Special thanks to Adis Jugo , Mustafa Toroman ☁ , Thomas Vochten , and the rest of the organizing team for this wonderful event 🙏 Already looking forward to the next one! #ECS2026 #PowerPlatform #PowerApps #Dataverse #PowerPages #Accessibility #AIBuilder #Community

Shipped a hobby project as OSS: OpenConveyor. I wanted an easy way to kick off AI agents from my phone, or wherever I happen to be, without giving them the run of my laptop or a long-lived service account. I spend many many years working with Kubernetes, and the primitives for this already exist: Jobs, NetworkPolicies, RBAC, projected Secret volumes, Pod Security Standards. The gap was a thin layer that composes them into something you can safely point an agent at. OpenConveyor is that layer. One CRD (Task) describes the agent, the prompt, and explicitly the secrets and hosts it can reach. The controller reconciles it into a hardened, short-lived Job and cleans up after itself. Anything not declared is denied. Timeouts are mandatory, so no forever running agent going haywire. Trigger it from a GitHub issue, Linear ticket, cron, claude code or copilot session with a /conveyor command, or Telegram bot. Any agent image. Any cluster. Apache 2.0, alpha state, written in Go. Links in the comments.