Last edited on August 18th, 2026.
Since the end of 2025 that automatically generating useful code can be unlocked by giving large language models (LLMs) the tools and resources to work in loops, specially if these agents have some operating system and access to the internet. But we should also be aware of a variety of risks ranging from the quality of the output itself to prompt injection from untrusted sources, not to mention a non-zero probability of having an agent wipe all your data.
Working with AI is subject to change almost every day, so this is simply where I stand during the summer of 2026 in an attempt to give coding agents full permission to run on a computer without overcomplicating things. Also, this is not necessarily my work setup but what I use for side projects:
- GPT 5.6 Sol High inside the Codex desktop app (that has been rebranded to just ChatGPT desktop)
- sole-purpose MacOS standard user with no access to external storage
- git remote access through deploy keys for each repository
- rclone backups to Backblaze through B2
- automated control of a real Chrome browser through Playwright
These current choices reflect two main requirements that seem non-trivial to put together. One is minimising the blast radius of what an agent can do on our system and the other is being able to share a real browser with an agent doing web or app development, so you can’t just stand up some container and SSH into it. Isolation in general is tricky if you assume constant remote code execution which can always leverage any vulnerabilities to break out of the sandbox. I am no expert, but I believe MacOS user separation is decent. Linux has a lot more options, ranging from the best solution if you know what you’re doing to something much worse than MacOS. Meanwhile, Windows is not as close to Linux as the other options, and its latest edition has affirmed it as the worst operating system in general.
Currently, my main gripe with this setup is having to go between admin and standard user accounts to perform updates or install any missing tool. MacOS is actually not that fast when switching users, even with the latest M4 Mini. MacOS also overloads your CPU by trying to monitor all the individual processes spawned by agents. As expected, Linux is still the end game for better performance and isolation if you can set it all up.
For specific tools, I appeal to what the agents already know, like bash or Playwright. The file system itself can work well as a memory system, and you still have SQLite as a classic alternative. For building desktop apps, I’ve been trying Wails as one of the webview-powered options, but I’m open to Tauri or even Electron. It’s nice to still leverage the same Playwright approach to keep validating the frontend.
About memory systems, I’m surprised at how far you can go with git and whatever md, yaml, html, or json files you want to put together. Agent skills are all built out of simple document formats and how LLMs already know bash pretty well. The challenge is to keep collapsing information into useful, portable, traceable, and repeatable units. Also, there’s no standard for authentication, and that’s where MCP servers may claim their stake.
Speaking of portability, I’m also looking at T3Code as a way to get unstuck from OpenAI while still using their models and harness. The thing with agent-powered projects is that they seem to generate lots of changes in a short amount of time, some good and some bad. How much of a beta tester do you really want to be? I look for projects that can establish trust in their engineering culture by how they communicate, make decisions, and discover best practices.
Finally, what have I built so far? Without using this article for self-promotion, I’m happy for now having:
- An automated website that tracks, translates and publishes daily information from sumo tournaments
- A desktop app to easily go through photos from my camera and select those I want to keep
- A skill that I can use in the ChatGPT mobile app to analyze and report on code repositores
- A personalized recommendation system for all the articles I keep saving to read later
- A web app that works offline in my phone to transform and clean copy-pasted links
- Completed a machine learning prize challenge with no formal knowledge of ML
- A desktop app to manage AI-generated images
- An improved version of this very same blog