Essays about game development, thinking and books
Coding agents often need to answer practical questions before making changes, such as: Which tests should be read before modifying this file? Which specifications govern this module? Which modules does this code import? Which artifacts are affected by this specification change? etc. In other words, the agent must discover all dependencies between files and add them to the context to complete the…
How Donna works. In February, I released Donna ( the original post ) — a CLI tool to run agent workflows as state machines and describe them in Markdown files. Since I was developing it simultaneously with learning how to work with agents , Donna came out as an overcomplicated, versatile monster. It had many features, all of which were useful, but in total, they restricted the developers'…
Rust is slowly but inexorably refactoring humanity's codebase. About 5 years ago, I was going through Rust documentation and decided that I didn't like Rust [ru] . Last year, I needed to prototype some game logic and chose Rust for that task, as I couldn't find anything better: Zig looked too immature, and C++ is getting more and more dead complicated with every standard. As a result, I wrote 10…
Life is like a hurricane here in Ham-burg… As you may know, I'm finishing my professional sabbatical. It is still unclear what I'll be doing next, however I've gathered a list of potential projects that are very interesting to me. I'm ready to discuss partnership around those projects right now. If you're an early-stage startup or are only just about to create one, I would fit well into the role…
Agents coding something (c) ChatGPT When I was summarizing the last year , I said that agents will be the main topic of 2026. For the last three months, I have been gradually figuring out how to use them for development, and now we can discuss the intermediate results. Formally speaking, coding agents are a subset of AI agents, but for simplicity, these terms will be used as synonyms in this text.…
An engineer demonstrates their work to the community (c) ChatGPT, Leonardo da Vinci Points of view on a product Hypothesis testing loop What to read, when, and why No instructions for engineering Engineering is science is engineering Scientific practices for engineering — public artifacts In the previous essay , we discussed how engineering and science are closely related, and therefore can borrow…
I've formulated my main conceptual issue with LLMs at this point, based on my personal experience. The problem is less noticeable in chats — it gets smoothed out by the continuous interaction with a human who steers and corrects the LLM. However, it becomes much more noticeable during vibe coding, or when you ask an LLM something that calls for a large abstract answer.
Looking for early adopters for my agent utility. I named the tool Donna — https://github.com/Tiendil/donna With it, your agents will generate state machines while executing state machines created by other state machines. Donna allows agents to perform hundreds of sequential operations without deviating from the specified execution flow. Branching, loops, nested calls, recursion — all possible.…
Right after I started working on the pricing for Feeds Fun users, I realized I should do it in a blog post: it's almost the same amount of work, it's the ideologically right thing to do, and it should be interesting. Anyway, I was going to write an RFC — the question is purely about publicity. I'm also taking the opportunity to conduct a retrospective on the project for myself. What is Feeds Fun…
Blog metrics for 2025. The New Year is near, so it's time to sum up the results of the year. Let me tell you what I was doing in 2025, how my plans for the past year went, and what my plans are for the coming year.