There's the core of a reasonable argument if you forget about using the word "consciousness."

Consider playing a game where some players are human and others are implemented by AIs. You might want to model your opponents' minds in a similar way? Someone playing chess or Go will need to consider what moves their opponent could make. If it's poker then you need to consider if they're bluffing. If you can make alliances, you'll need to consider how far other players can be trusted.

You can put scare quotes around words like "intention" when discussing AI players if you like, but it might not matter when you're figuring out how to interact with them.

When a game is complicated enough, an AI player is going to need memory (likely implemented by the harness) and it will have to model the other players too in order to coordinate or compete with them.

This doesn't happen when I use a coding agent in a normal way to write code, but I haven't played around with persistent agent stuff like OpenClaw or Gas Town. They seem to be making their agents act in an unnecessarily complicated social environment, which seems to lead to weird situations?