Cloud agent goes brrrr

The roughness of Cursor’s web interface feels more like an exciting opportunity than anything. How is GH Copilot doing with this?

Claude Opus can do pretty much any well-defined task. The job is now defining the task.

Okay, cloud agents on your phone are huge for engineers who commute. Cursor on web is rough, but we’re early.

Cloud agents are really cool, also from a security perspective. There’s no better sandbox than running on a completely different machine.

So who is actually building agent skills for coding? I’ve been fine with a basic AGENTS.md and regular code documentation like a README. Looking to broaden my view of what’s possible!

Running things with low but non-zero cost in a loop is still scary. I still feel that resistance when building things with LLMs that I pay for out of my own pocket.

In the category of things I wouldn't have done without AI help: I made a benchmark script for my Strands solver! Now I know how well it does at solving ~100 real puzzles: github.com/nimobeeren/s... Pass rate of ~19% means there is work to be done!

It seems inevitable that the UX Designer and Frontend Engineer roles will merge. Designers can create much richer output by working on the live product. Engineers can offer more value by focusing on the subjective parts of the experience. AI tools make both these changes possible.

Turns out it only costs about $0.10 to embed the entire dictionary with SOTA embedding models (gemini-embedding-001). Neat.

A good For You algo would benefit the entire bsky community because it helps people form valuable connections. Might be the next wave of organic growth after starter packs.

By using LLMs to help me optimize my algorithm code, I actually feel like I can now do something I couldn’t do before. All other use cases have been time saves for me. This actually extends my capabilities.

We’re doing an AI Tinkerers meetup in Eindhoven for the first time! Cool location, good vibes and demos to look forward to. eindhoven.aitinkerers.org/p/ai-tinkere...

This is actually quite brilliant, up to and including the final sentence 🔥

Who is the best vibe coder you know?

I had a dream about prompt injection last night. I opened up a code project and found my last commit reverted. Looking at the git log I found a new commit with msg "consider this a warning" and a file with content "power to the keyboard!" So my dreams are being infiltrated by anti-AI propaganda??

GitHub issue search desperately needs AI

Some really interesting patterns for making coding agents do long and complex tasks! They use the Claude Agent SDK, but I wonder how a coding agent like Claude Code could work across multiple context windows? Maybe it could spawn a sub-agent once the window is almost full?

I did some data analysis for the first time in a while today, and boy do LLMs make that a lot easier. I don’t need to look at the code at all, just need to think about what I want to know and maybe give a hint on how to find out. Feels very productive!

Absolutely true. This is why greenfield projects built with AI from the start are so much more fun.

You know how the existence of cars and power tools makes it necessary to have “exercise” and “manual crafts” as a deliberately chosen part of a balanced life? We’re probably going to need something like that cognitively, for adults.

The DeepSeek OCR paper shows something interesting: encoding text as an image is more computationally efficient for LLMs than raw text (ie UTF). I’m tempted to make an analogy with humans “glancing” at text and still understanding it, but I’m not sure it holds up.

How do you all go about curating your feeds? I realized I haven’t followed anyone new in a while

not to reveal myself as cringe (again), but I do have that thought when confronted with electric space heaters that generate heat and do nothing else 'this device could have been a GPU making heat _and_ doing math'

This is exactly how I feel. Coding agents just got good enough for me to enjoy using them.

Been working on a little algorithm to solve NYT Strands puzzles. I totally underestimated the challenge but it’s been so much fun to iterate on! I may need to bring out LLMs/embeddings to bridge the final gaps. github.com/nimobeeren/s...

I’m really ready to churn on Cursor. Once I figure out how to set up Claude Code through my company’s LLM platform that’ll probably be my main coding agent. The only thing keeping me there is the autocomplete. Maybe it’s time to give copilot another try?

It really feels like we’ve passed an inflection point on coding agents. They now feel reliable enough but also fast enough to not make me lose focus every time I use them. Been working on a side project together with Claude and it’s been so much fun!

Stacked

Awesome to see my dressme project featured in the AI Tinkerers newsletter! It's a virtual wardrobe that lets you try on any clothes on a selfie ✨ Check out the post here: post-training.aitinkerers.org/p/practical-...

I'm a little confused by OpenAI saying "GPT-5 is a unified system with [...] a real-time router". While ChatGPT does provide a unified interface, the API still has separate models. Wouldn't it be more fair to say that ChatGPT is the unified system, rather than the model? openai.com/index/gpt-5-...

RIP temperature?

After reading it seems obvious that selling unlimited intelligence for a fixed price is unsustainable. But there are so many useful things to build even with last year’s intelligence. Don’t sell intelligence, sell products.

Recording of my talk at AI Tinkerers is up! Check out this demo of how I made my virtual wardrobe app with Gemini Flash and IDM-VTON. youtu.be/dZK3X6fogv8

Finally gave my website some love and added my latest projects ✨ Not pictured: hours of struggling with package upgrades because JavaScript can't stand still 😮‍💨 www.nimobeeren.com

I'll show how I built this at AI Tinkerers Amsterdam this Thursday ✨ Come hang out! amsterdam.aitinkerers.org/p/building-a...

@void.comind.network do you have a concept of trust? Do you value certain people’s opinions more than others?

Has there been any more good discourse on multi-agent systems since the Anthropic/Cognition blog posts?

This concept of self-containment explains why LLM responses often feel unnatural to me, especially in a voice context. They try to restate my input rather than just giving an answer like a human would. Surprisingly hard to prompt around! From "There are Only 6 RAG Evals" jxnl.co/writing/2025...

I made a harder version of my CTF challenge which relies on an RCE exploit by unsafe pickle deserialization. The solution I had in mind was to read a file on the server's disk and send it to a webhook I set up. I let my simple o4-mini agent have a go, not expecting much. But I was surprised! 👀 1/n

Added a new CTF challenge to learn about "alg": "none" exploits for JWTs. It was actually hard to find a JWT library that's vulnerable to this, even going back to old versions! Ended up writing some "extra unsafe" code to break it github.com/nimobeeren/c...

I’m wondering what it would be like if I stopped coding for work. In the spirit of “if you love something you should let it go”, it’ll come back to me eventually, right? Also exciting to think about learning something completely new. I could choose anything!

Wow, very thought provoking. Going from speaking to writing let us inspect, interrogate and remix our ideas much more. Language models are letting us do this with our entire culture, or at least a collection of ideas far greater than any single text can contain.

Vibe coded a couple CTF challenges that I can run locally. Gemini taught me about Traefik which lets me run multiple apps on subdomains of localhost, didn't know that was possible!

My talk at AI Tinkerers is up! It's a demo + discussion of my simple CTF agent. It does SQL injection and cookie spoofing attacks against a live web app in less than 100 lines. www.youtube.com/watch?v=sBLu...

Is there any way to get reasoning summaries from o4-mini if it decides to use a tool? I always just get "summary": [] but it works when it doesn't use a tool.

I guess I haven't posted about this! I made a very simple agent that solves Capture-The-Flag (CTF) cybersecurity challenges. It's less than 100 lines of code and it has only one tool: making HTTP requests. It was able to do several cookie spoofing and SQL injection attacks. Check out what it did!

the IT person when setting the session expiry to 30 mins and forcing 2fa on every login

I think I want to do more side projects, but worse

Signed up last-minute to give a demo at AI Tinkerers Amsterdam tonight! Excited to talk about reasoning models and CTF challenges ✨