Our Collective Chaos group is starting a book study today in Liturgies of the Wild by Martin Shaw. I don't know that much about it yet because I am just sitting down to read it now, but it's something about the benefits of myth, why we have largely lost it in our modern cultures, and how to get it back. If you are interested in joining let me know and I'll get you an invite!

Today I realized I can no longer do long division by hand.

One of the tasks my local AI assistant does every morning is to sift through a bunch of news sites and pull out articles that fit some conditions I've set. After we had a hundred year flood in my city last month I had it add the local city government alerts page to the list of sources. But there's a bug where it seems to sometimes just pick a random page from the government website that isn't news or an alert. I haven't tried to fix it yet because I'm learning a lot about my local government in the process. TIL that the drainage board is meeting this week. I have never heard of a drainage board before but I guess they approve plans for drains and ditches. It has been a nice way to grow in appreciation for all the hidden work it takes to have a functioning city.

Also someone asked me recently what I have been using my personal local agent for. Here's what I wrote: Hermes w/ Qwen 3.6-27-a3b on an old gaming pc in my house. Looking through my threads over the past couple months I have used it for: - morning email with local + global news articles (attempting to filter out FUD and clickbait), weather - afternoon email with plant care todos - learning about new topics. Recently: how bonds work, how to test flower bed drainage, meal planning ideas - I have a product finder skill where I have it do web research when I want to buy something. Recently: electric toothbrush, journaling app, microphone - Recently I have started asking it to do security audits of open source software before I use it. I don't really trust it but it often pulls up real red flags. - arranging travel plans. It mostly does research on flights/lodging/food - watching and notifying me when websites change. Recently: Dune 3 movie tickets, app release fixing a bug I reported.

Almost two months after canceling my Cursor subscription and it is going much better than I thought. My AI development stack has been 100% on open weight models for two months, and my personal AI assistant is still going strong on 100% self-hosted models. I haven't used Claude, Codex, Cursor, Grok, or Gemini basically at all the past two months (may have done some website design with Claude one day? Can't remember) and the surprising part is that I don't miss them at all. - I don't miss the subsidized subscription games that Anthropic is always changing the rules on. - I don't miss the frontier intelligence - the open models feel 95% as good. - I don't miss the constant suspicion that these companies (especially Anthropic) are making their models dumber or slower weeks after release. - I don't miss getting kicked off a model I rely on because a newer version is out (Opus 4.6 was the GOAT, change my mind) - I don't miss every version of my harness introducing new bugs and changing how things work. I've been running mostly on deepseek-v4-flash with pi harness, and a homegrown orchestration system using a combo of beads, fragua, herdr, and a custom web UI I call mirador. View quoted note →

@verbiricha I've been using fragua for a couple weeks now and I quite like it. Some feature requests: - i want tool steps to have structured output (for example i have a tool to compute branch name). Right now only llm steps can provide structured output? I've hacked around this by having an llm call my tool, but it's not ideal. - I wish the webui had some more contextual actions (i.e. approve/decline changes) on the run's detail page - I wish there were more timestamps in the run list and detail pages. - I want some control over worktree creation. I have a workflow that I use for feature work or if I pass it a PR url it looks at my review and addresses it. This leads to worktree collision where sometimes the initial run for that branch left tree dirty or the branch checked out and fragua fails to account for this. I've worked around it temporarily by creating a new branch for every run. I'm not asking you to implement all this. Just letting you know in case it dovetails with feedback from other users. I am not using it as intended where I want to apply the changes to my local git worktree at the end of a run. I am running fragua on a remote sandboxed machine and most of my workflows open or update a PR that I review elsewhere. Which makes the Approve/Discard step at the end of the run kind of unecessary. I'm thinking about building my own webui that is beads-aware but still uses fragua to drive workflows under the hood.

I canceled my Cursor subscription today. It has felt for a while that they stopped building for the way I want to work with AI. Then the bugs keep getting worse, and the SpaceX acquisition doesn't give me a lot of hope that things will get better. I spent yesterday using VSCodium and my homebrew agent orchestration system in tmux. I also tried emdash (https://emdash.ai/), and retried Archon (https://github.com/coleam00/Archon). I'm still feeling like small agent swarms with a mixture of models that pass work between themselves is the sovereign AI stack of the future. The manager of the swarm can paper over a lot of the small issues with LLMs. Plug in whatever cheap source of tokens you can find, mix in local models if you want, your interface stays the same. Put it all in a decent sandbox and dangerously skip permissions.