Mike Mehanig

What are we gaming with 8Gb??? Windows bootloader?

draken.mehanig.com If you run multiple agents 27/4 it’s hard to keep track what happened, when and also why. Draken builds searchable knowledge from your agents actions and you talk with it, similar to “Daily Sync” Some bugs there still but looks useful already

I was using Expo before, to build my iOS apps. I now switched to just running OSX-KVM on local Fedora server, and the next step is to completelly migrate to native code. Dropping dependencies was never been easier...

Something weird happening with apple security. I was using account1, and was able to get 2fa for account2 then approved it via phone from account1. What The Fuck?

Do we even need CI actually? For agentic PRs, assuming no security vulnerabilities and no agent misbehaving, it’s mostly waste of time… old standard of ci/cd is obsolete and should be rethought.

Self hosted is the way to innovate

Half of this list only applies to greenfield projects. “Generate me 20 versions of deployment for service” ?

Corrected: “If you're in software development (…) you sort of have to decide, am I retiring now or am I re-skilling?”…

I can finally see experience like: “drop media, drop script, create video for me” working cheap and reliable!

Soon Gifgun.online will have its own agent 🤔

I really hate that “Login with ChatGPT” capability requires approval from OpenAI… this is literally the last piece of the puzzle to run Pi agent in browser sandbox persistently and without headaches…

Finally, my dishwasher is Agentic!

I wish there was a way to say to the browser - DON'T BOTHER ME WITH CORS... This is the only thing stopping gifgun.online to have it's own In-Browser AI agent, that doesn't need any infra hosting or extra server relays. (BTW this experiment was based on wasmer.io and it's amazing piece of technology)

Somehow I think this is the new meta, that’s being build at gifgun.online Instead of blindly allowing agents to click around web pages, page publishes local bridge and local agents can interact much faster and more reliably, while also spending less tokens.

TIL: orange cloud has 100% sla

Evals are underrated, if your repo has skills or does LLM calls, your evals should be solid.

I absolutely appreciate that there is separate security team that manages our private registry for npm packages. I can sleep pretty well at night.

Someone reached out to me about my abandoned project: Draken draken.mehanig.com While I abandoned it pretty quickly, it’s time to reflect on why I abandoned it. Idea was relatively straightforward: every prompt via draken will spawn separate docker container to work on this promp, and duplicate claude session files that are mapped inside this docker. Claude will run in docker on copy of files there and you supposed to have more control and visibility into claude code changes. But after using it myself for a few days, the verdict became clear, what I build was not useful. With current state of LLMs and agent harnesses, I rarely need to review or re prompt from a specific checkpoint. It’s much faster to just stay in a single session and work iteratively, then try to be “smart” and branch from a session. Some things just don’t need to be built, and I am very happy that Agentic engendering allows such quick iteration from idea to MVP. Time well saved

I loved pycon years ago, it’s sad to see that it’s getting worse :(

Instagram feed: I traveled here; I ate this; X feed: i WrOte sKiLL X is superior.