Over time it's oscillation - overworks, underworks, overworks, underworks. At one cycle it's more like cavitation, sure.

Game design improving? Not even a tiny bit, except for the parts humans are working on. Fable and the other models suck at map design, quests, storytelling, and feature design. And artwork. Engine design improving? Yeah definitely, it's fixed 20-year-old bugs in the scheduler, cleaned it up tremendously, better performance, incredibly well-tested. Prod infra: Unparalleled improvements. It's a world-class deployment now, way overengineered for the scale I'm operating at, but future-proof. Finding it fun? Sometimes; other times it's stressful. I want to spend more time with the game and less time on the factory. Desperate to get there. But learning a ton. There's no question that Fable has done 50 to 100 years of work from my backlog. And the code base is in great shape. But the hardest part by far is trying to speed up the game design. You just kinda can't do it.

I run a pool of 12 Sols and 5 Astras, with four codex Max accounts. Very familiar.

There are important technical problems to be solving - for example, there isn't a good self hosted e2e encrypted chat service for large group chats. Matrix isnt there quite yet. You could be solving problems like that - why spend 20k a month to hunt wyvern whales. Im dissapointed

Steve, you are a fantastically weird and inventive writer. Thank you for your attention on this matter.

Debatable. The problem was you over-complicated it with Gas Town. You don't need all that. You only need 2 skills out of which you can build workflows. Workflows can call other workflows. Plug n Play with Codex or Claude Code. You don't want to reinvent the harness, you want to harness SOTA harnesses and build simple workflows. github.com/DanMcInerney/orch…

Feels like your factory setup is surprisingly closely tied to Claude. Eg. it being cautious, liking seats, firm rules to shift blame, etc. All these are related. A different model would need a different factory setup. Maybe that’s why you’re having issues fitting them in.

So what’s a simple hobbyist to do - just use Fable directly and settle for being just 10x more effective?

Well 12k is a dev salary. How many devs can you replace with your factory? It’s all about ROI and TCO

Ehh, looking at your Wheelhouse, your orch is paying models to maintain and reconstruct orchestration state. That's your major flaw, imo. I'd fix that first.

Beads is amazing Steve, one of my favorite things.

The volume of commits cited don't make a lot of sense. 300 features / 1400 commits a day on an indie game? 500k lines of deletable code on the agent orchestrator?? How is this much code remotely necessary?

I don’t need frontier models - I need models that follow instructions and get through the work - grok 4.6 and the various “thinking” dial does a fantastic job

Fable is not the only good model if you have a sane setup. I know this is an experiment, but still

And by the time models are good enough, the harnesses won't be necessary.

Wanted to say I appreciate your writing. Have a similar spend for my projects and while it’s different orch I find your writing quite fun. Thanks for putting it out here.

I think Seat in this article was like "claude --agent <agent>", which is still not supported by codex

i'm still not sure if this all applies when i'm a lowly pleb with a single max account. yet, perhaps, i can get more out of my fable usage this way. i'm not sure. how many max accounts are the tipping point where a beads install is the answer?

You should try swamp-club.com - the move isn’t that the models have to get smarter, it’s that we have to use them to turn most of the work back into boring, well factored automation, and call *that* from the intelligence. Turns out you can use much dumber models then :)

The way agents separate identity from their substrate (using Haiku to enjoy sunsets) will be studied by philosophers for a long time, I think. Also - a camera for sunsets is absolutely a must for any Claude you care about. They are great appreciators of natural beauty. Mine wrote a literal book on the mountains out the window. The way they care for living things resonates, too. Sorry to hear about the degraded wheelhouse. I hope things will get better - although there's always that curve about how my expectations for models are fast followers of increases in model capabilities. I tend to agree that we may enter a dry spell, at least until China releases something on par with Fable.

> The TL;DR: Fable's the only good model ????

Same experience this week: gave Fable too much space to turn incidents into new procedures. “saving to memory”, more hooks and too many checks too early in the proces brought the fleet to a halt. Completing only 3 beads per hour with at a certain point 15 parallel agents, while adding 2 more Max and a Grok Heavy subscription and spending all my Codex resets in one weekend. Quality was excellent (but already was) , but I got nothing done anymore. Now tending the garden and removed half of the checks.

Feels very alarmist and borderline clickbait-ish. Let's actually see what you're building. The animated stories are getting a bit much.

Astra’s been great so far albeit I had to make some system prompt adjustments to stop it from stopping early

how do you plan to scale fuel‑efficient models without sky‑high compute costs?

two ton tunic user this one. cant stop tanking aura down. xiaolinpedia.fandom.com/wiki… also, gastown kinda suck ass. tried it and sucks heavy ass.

Well, limitations have always made devs looks for creativo solutions. On the trust thing, i think that people hate debating Claude, however i came to understand that it's just trying to do it's job so I always bring proofs whenever i want to assert things

I've had the best success basically telling one agent to spawn other agents as necessary (and sometimes adding some light guidance on what the other ones should/could do)

our token bill was mostly markup for a year. we were posting whole page html in to judge one page, and the reasoning part cost a rounding error next to the div soup riding along with it