In multi-agent game generation, I spend a lot of time thinking about what gets passed between steps. Requirements, layouts, assets and code can come from different agents, but they need to refer to the same version of the game.
An asset might have been replaced while the code still uses its old address. A step might have been retried while the next one kept its earlier output. To investigate either case, I need to see the files each step received, what it produced and whether it finished successfully.
I prefer keeping versions and completion states explicit. Conversation history helps with context, but file references and current versions should be easy to look up directly.
I don’t have a fixed preference for how many agents to use. Some tasks are useful to check and retry separately; others are simpler to keep together. Each extra handoff adds waiting and coordination.