For AI game generation, the division of work seems fairly obvious. One agent handles requirements, another the interface, another the assets, and another the code. It looks tidy in a diagram.
Once they’re connected, I have more specific questions. Did the coding step receive the double-jump requirement? Is it using the latest asset URL? If an earlier step was retried, which version of its output is everyone working from?
Those questions don’t stand out much in a single conversation. They matter when someone needs to use the product repeatedly.
I tend to start with the inputs and outputs of each step. What has been done? What is still missing? I’d like to be able to check that, rather than ask the next model to interpret a long message saying the work is complete.
The workflow also needs to keep track of its state. When something fails halfway through, I want to know where it stopped and whether we can resume there. Starting over can throw away the parts that were already fine.
I don’t have a preferred number of agents. Smaller tasks can be easier to inspect, but they also introduce more waiting, retries and handoffs. If one step handles the work reliably, I’m usually in no hurry to split it again.
When I see a complicated agent diagram now, I mostly want to click one of the connecting lines and see what’s actually being passed along it.