This infographic is just there to pull people into reading what I am asking.
Is there some kind of a meta when you develop games with AI agents?
Like which engine to use? Which 3D generation tools etc.?
Lets say Gemini 4 at the current state is really good, but not the absolute best: Better than Sol, but worse than Opus 5.5, Fable and Astra. And lets say they can compensate that with cheap pricing.
Would the management of Google release such a model or will they delay until they can show a model with benchmarks where they are the #1 in almost every other frontier model?
I used muse spark through OMP which ran 4 parallel subagent to generate a 2D game map (a lot of visual understanding), then after it was finished and I just asked to tell why it took so long, I got an error:
>401 This account has found to be committing fraud or is in breach of terms of services and has been blocked. If you
> believe this is a mistake please send an email to help@anoma.ly.
> This account has found to be committing fraud or is in breach of terms of services and has been blocked. If you believe
> this is a mistake please send an email to help@anoma.ly. (type=AuthError)
What was the possible reason? And how long will it take until opencode activates the account again?
Since Israel showed the world how cruel they are, they are now trying to distract from their crimes by blaming Islam/Muslims for everything. The idea is to create hate towards islam/muslims to justify their war crimes: "Wuaaat? Are you on the side of these barbaric LGBTQ+ hating people????"
So fellow Pakistani Gen Z folks, take care!
I subscribed opencode go, and use it in Hermes agent and OpenChamber. I got an email from opencode just now like this. Will it has any unkown limit on my usage? If I try to use opencode go in other harness, like pi or Zcode, will they fail after 09/06?
Opencode Go offers Muse Spark 1.3 Contributor for free.
Would be cool if Command Code could also offer this.
Since it's the sane model, just with more post training. It consumes the same amount of VRAM on GPUs, so for the providers it would mean a free upgrade.
wanted to know if it makes a difference for easy tasks or can I just stick on 4.6?
If you use Grok 4.6 for coding, where do you get more usage/tokens:
Cursor 20$
Supergrok 30$
I feel like everyone is saying something else.
I like LLM subscriptions where I can use the API for third party tools like Pi, Opencode, Hermes etc.
But with Gemini it is not possible, you can only have Antigravity running on a VPS which is called from another AI agent.
Is this something that makes Gemini modles out of scope for vibe coding? Or do you accept it and use their harness?
I want to use Grok 4.6 but am not sure about the amount of token I would get. Where do I get more and what is the ratio?
I currently have the Claude Pro subscription, but Grok 4.6 has been released, which looks really strong.
Since both Opus and Grok are good enough for me, my question is: where do I get more usage? Claude Pro or cursor?
There is 20$ allowance for Deepseek V4 Pro
And 70$ for Luna
When I consume half of the V4 Pro allowance (10 out of 20), do I then have 60$ gor luna left or just 35 (half of the 70$ budget)
What do you think of it?
https://commandcode.ai/pricing
It seems like it has for some models more than Opencode Go, in some models less.
But what is the overall opinion?
Is there a Skill, a harness a tool or something that abstracts away all the stress of thinking of how to create a good loop for long running multi step tasks?
My current approach is still in a way that I ask an intelligent model to create the architecture and a long list of TODOs. Then I start a new session where I say "go through the points in the TODO and solve each one after another by using a subagent".
But I would be happy if there was an abstraction layer that I can just say "create a Mario Kart clone" and it just does it without trying to do everything in one huge run.
Is there something? I am open to various agents like hernes, opencode, codex etc.
EDIT:
To maybe clarify my question:
So one thing is to have a loop that goes through all existing TODO items. That is pretty simple.
But the challenge for me is: How do I create a loop, where the AI agents creates new TODO items on its own. When I say "create Mario Kart for me", I do not expect that it should create all 100 TODO items at once. I would expecte a loop where an agent looks at what is finished already, what is missing and depending on that creates new TODO items on its own.
I am impressed that Qwen is able to finish long horizon tasks by running for several days. But I wonder: Is that feature of running for so long not something the harness gives? If I would ask hermes agent using Qwen 3.8 MAX "create a World of Warcraft clone for me" I am sure it would immediately start coding some stuff, fill up its context windows and after 150 iterations (limit of hermes) if would stop with some half-baked broken stuff.
So whats the way to enable long running tasks?