I just realized it's been almost a year since I opened a proper IDE. I don't say this is good. I don't say this is bad. I just think that the way we work changes. No matter if we accept it or are still in denial.

One of the core parts of my AI coding setup are domain specific skills that define and describe each of our main domain concepts. Both technically and non-technically. That's when I can run prompts like "/investigate X's DEE score and explain anomalies"

I probably wouldn’t have considered this possible six months ago, but I’m actively looking into moving away from Claude Code! And I don’t mean just for personal projects, but for all Atherio-related work as well. There are quite a few reasons for this: 1. After Opus 4.8, every single new model release felt like a fight. Models fighting against rules, documentation, skills, and every type of instruction. Opus 5 was probably the worst. Because of this, every single task consumes way more tokens than it did with Opus 4.8, for instance. Meanwhile, token prices for the newer models have obviously increased. 2. I’m quite sure flat-rate subscriptions will soon disappear in one way or another. Every business needs to get accustomed to sizing its agentic workflows around API pricing. The sooner, the better. 3. Related to the previous point, I feel we need to get to a place where we need much more flexibility to choose the model we want to use for specific tasks. Being confined to just Anthropic’s models is not optimal. Models like GLM 5.3, Qwen 3.7, or even DeepSeek 4.1 Flash have proven that they can perform many day-to-day tasks on par with Opus models, at a fraction of the cost. 4. Going through the Cyber Essentials certification recently also showed that Anthropic’s setup caused us the most problems. That’s why we’ve already moved all non-engineering work to ChatGPT. This switch isn’t something that can happen overnight, since I still need to validate a lot of things and do quite a bit of work. But I’m now actively going through that process, with the goal of cancelling our Claude Code subscription by October 26th. Curious to hear what other Claude Code power users think. If you’re considering switching, why? If you’re not, why not?

The first day experimenting with OpenCode in our production system was better than expected. - Validated all our monitoring workflows - Validated investigation workflows - Ran a few implementations of already planned work (quite simple work though) How much did it cost us? $0!

One of the good things of having developed my own tool to track my agentic coding is that now I have clear data on how many cached, input, output tokens I use per month and forecast what that would mean in terms of costs if paid at API prices for different models like Kimi, GLM, DeepSeak, Grok

A cheap or even free LLM that actually uses your skills, rules and instructions performs just as good as a SOTA model that constantly refuses to use them. At least for coding

For the fist time today I installed @opencode and opened the Atherio project. Was very surprised that even with the free Big Pickle model I could perform some work that used my domain specific skills, subagents and rules without any configuration needed.

Let me tell you how we're using AI in sales. We have a bunch of agents and scheduled tasks that find us propsects, categorize them, create and move needles around Hubspot. But the outreach is always done by us humans. We never let AI agents send emails or any form of communication

So, the past 2 weeks I've been mostly traveling and talking to customers, investors, joined a bunch of business networking events. Clearly didn't concentrate too much on tech. So, can anybody tell me the summary of why Jev became such a hype?

we always need to adapt our agentic coding workflows to the value we generate with the performed work. If you spend 10k on tokens, that work better generate at least 20k in revenue.

Microsoft recently ported the GitHub Copilot runtime from TypeScript to Rust in roughly 4 months. The entire port was mostly guided by one single engineer and generated a token cost of around $120k. Here are few of my thoughts through my CTO lens. 1. Why Rust and not C#? With AOT, C# was theoretically a possible choice. Yet they decided to use Rust. I'm quite sure this was technically the right choice. But this puts C# and .NET in a tough spot. On the one hand Microsoft hypes the huge performance improvements and the AOT capabilities. Yet for their own products that have to deliver at high scale they chose something else. 2. Mostly one engineer took care of the entire port that was executed mostly by coding agents. However, it's important to note that the engineer doing this was Stephen Toub, which is an absolute god when it comes to performant code and low level programming. So, it was not a random "X expert." Which leads me to put this into the "don't try this at home" category. 3. The $120k cost seems something out of this world for most small and mid-size companies. When it comes to token costs, I think it's always a matter of financial context. I'm quite sure that this port generates enough ROI to make the token cost negligible. The learning point here is that we always need to adapt our agentic coding workflows to the value we generate with the performed work.

Last week it was "strategy day" for Atherio, so we met face to face to reflect on what we achieved so far and what's our game plan for the next 3 months. There were some novelties this time as it wasn't just me and Ben. Starting August we also have Will on board. We're really fortunate to have him joined our team. And, yes, we kept the tradition to start the day with a run. This time the run brought us to the Emirates Stadium in London, which is quite cool. I also swear we don't have this in our hiring policy, but Will is by far the better runner among us. The discussions we had were very insightful and now we have a precise gameplan for the last 3 months of the year and we're all excited to execute it. More to come!

I have a big challenge to take seriously tech influencers whose every second post is a promotional one.

Today we got together in person all 3 of usvat Atherio to look back on what we achieved so far and also plan our execution for the final quarter. One fact: we made 84k revenue this year. Is it okay for a 10 months startup? Is it bad?

Changed scenery for today's run

While setting up my new dev machine it hit me that I didn't install neiter JetBrains Rider, nor Visual Studio or PyCharm. And it hit me that it's over 1 year since I last used a full-blown IDE. Am I the only one? Who's still using those traditional IDEs and why?

Wondering who at Apple had the idea to call their default mouse scrolling "natural scrolling." What on earth is natural about that? 🤣

"I'll never swtich to anything Apple" was something I said all the time. However, for the sake of some amazing stuff we plan to do at Atherio the next quarter I needed to sacrifice my principles.

Dario Amodei’s latest essay raises challenges we absolutely need to take seriously. As AI capabilities advance, questions around control, alignment, ethics and ultimately what we want these systems to become are no longer theoretical. As technologists, we have a responsibility to engage with them. But I also think we should read the essay in its broader context. Open-weight models have improved dramatically. For many engineering workloads, organisations no longer need to rely exclusively on frontier labs. That changes the economics of AI, particularly for companies whose competitive advantage depends on being the gatekeeper to the most capable models. And this is happening while enormous financial stakes surround companies such as Anthropic and OpenAI. That doesn’t make the safety argument wrong. But regulation can serve two purposes at the same time: reducing genuine systemic risk and reinforcing the competitive moat of incumbents. As CTOs and engineering leaders, I think we should be capable of holding both ideas at once. AI safety deserves serious attention. So do the incentives of the people defining what “safe” should mean.

Why do people say macOs is "intuitive"? It's literally the least intuitive piece of software I ever laid my hands on