It's pretty amazing to write your own agent BTW. I've got a zero-dependency all-in-one-file agent harness I wrote myself. I use it all the time now because I can get it from anywhere and I can know EXACTLY what it'll do (as much as you can with any model), what it's been told vs not. Using it as a harness for models I'm hosting myself makes me feel like some kind of LLM homesteader: it's a set of tools I'll always have that will only change as much as I want it to change.

My inspiration was the mayor idea from Gastown, plus wanting to formalize the informal workflow I used with agents and Jira at $dayjob.

As for the Orchestrator, it's pretty simple. In essence, it's like "Jira/Trello/Kanban on autopilot". Work items have states, a state machine defines how those work items transition between states, states are todo, in progress, retrying, reviewing, code reviewing, done. work items also have connections, allowing the LLMs to specify a dependency graph, and the dependency graph informs the dispatch order/parallelism, as well as when branches have to be merged. I talk to the steward, the steward has tool calls for interacting with all the data, and the orchestrator auto-dispatches all the work that comes in. I can generate work as fast as I can describe it to the steward, and that's usually the bottleneck.

So far I haven't had to deal with "how do you get the LLM to re-organize the work mid flight due to a worker finding something not accounted for by the planning", but I assume it'll come soon. The most complicated digraph I've tossed at it was 9 items and 4 layers deep. The kind of work I've given it hasn't been scoped large enough yet, so we'll see how it tackles that.

I'm still using Claude at work (they're the only approved provider), but wow are the smaller models starting to SMOKE the big ones. At this point, all I'd consider paying out of my own pocket for is the lowest-limit Anthropic/GPT plan to get a big model as the Steward, but I wouldn't pay for ANY of the Anthropic models as the workers who do all the work. And as time passes, I don't know if I'd even do that; the open models are serving SO well.

But theoretically, the output of every LLM is seed-driven (or could be if you wrote the software to isolate it) just like any computer software. It's just none of the software written (even llama.cpp AFAIK) chooses to support stable-seeding due to the changes in stuff like CPU/Vulkan/CUDA/Metal differences making it difficult to make consistent.

They could though! Hopefully one day someone implements it into the mainstream LLM-engine software and it gets exposed in the APIs serving the models. It'd do a lot to show folks the "internals" of these models.

Direct democracy is cool, but also impractical. I do not want to vote on every counties appropriations for road maintenance. So what's a level of direct democracy that's "good enough"? How do we make sure we're directly voting in things relevant to our lives? What if "relevant to our lives" is unrelated to our geographic location and is very interests based? If anyone can vote for anything, but most folks don't ever vote for most things, how do you prevent brigading of votes via coordination by groups who see that their group alone can swing what would be a small local vote whatever way they want by virtue of sheer numbers? How do you prevent trolls from going through every vote and just voting no on every "community center paper-and-ink budget" across the entire country?

There are so many questions I have about direct democracy systems! Do you have more information?

Does the harness parse the contents of CLAUDE.md, see the @AGENTS.md and then inject the contents of AGENTS.md into that file with a note saying "hey Claude-the-llm, here's the contents of AGENTS.md per an @reference"? Does the harness parse and inject but WITHOUT any ceremony telling claude-the-llm that what it's looking at is from another file? Or does the harness do nothing and Claude-the-llm is merely trained to see @references and treat those as "I should go read that file"? Which is it? If it's #1 and #3, then on long contexts/after compaction, then the LLM is still pretty likely to lose the fact that it read that file and re-read it again. If it's #2, the LLM could have the same problem as with the symlinks.

How does it work? The Anropic docs certainly don't say. So what's the best choice? I guess it's magic. For folks like me who care though, I guess there's always Pi or Opencode (or even, writing your own harness [which is shockingly easy if you want it reliable without all the fanciest features]).

Roads used to be for people and wagons, till cars showed up and kicked the people off. Now delivery bots are trying to do the same thing, kick the humans on foot off the sidewalks.

That's not _nothing_, but it's pretty close to nothing, and for the prosumer crowd it edges towards "just gets in the way".

> "Quality is like running edge models from 8-12 months ago"

Don't expect Opus, expect more like Haiku. If you micromanage it, you'll get great results. If you want it to be a human in a box, it'll flounder.

If we actually limit them to "only accepts tiny ultra well defined problems and ultra well defined outputs" then theycease being a $10T/year idea and become a merely $10B/year idea.

Thus, it is not exactly popular at the moment.

Take Okta login for example. Okta wants to offer big hyper-secure customers an option of "if you want, we can run our system in your cloud/data-center/whatever". To support that kind of system, you go to to the https://login.okta.com/ page and enter your email, JUST your email. Okta uses that to look up which customer tenant you belong to, then sends you to customer.okta.com where you enter your password. This way, the password only goes through infra owned by big-customer.

Okta then just builds everything with his indirection so they can move customers to it.