I asked qwen3.6-36b to make a zombie-like game on the web. I told it not to work alone, but to summon sub-agents. First, find out about zombies and then summon more sub-agents - coder, reviewer, team leader - to form a team and calculate the game production, making phases, and resetting the context each time a phase was made. /gol I used this to make it all at once.
It took qwen3.6-35b 4 hours and 30 minutes to make it, but it wasn't even a zombie style.
I ordered the qwen3.5-122b model with a prompt similar to that of qwen3.6.
And it completed at an incredibly fast speed.
When I asked it to make a zombie game, it made a snake game.
I told it to summon sub-agents and manage the context, but it just did everything alone.
Both of them didn't understand. I thought qwen3.6 had amazing coding performance, but it would be hard for it to interpret and recombine. But even though qwen3.5 is 122b, it was pathetic.
Still, the fact that it worked alone makes me wonder if this is just a parameter difference?
The game itself is cleaner with qwen3.6, but that's the end of it.
If I properly fill the prompt and harness and give it more tools, I think it will be able to make something even better.
