Dimensional Double Shift was designed as a cooperative VR game. Bringing the full experience to solo players meant replacing up to three humans with generative AI teammates that can follow a fast-changing game, coordinate with each other, and act quickly enough that the player never feels like they are waiting for inference. That creates a difficult production problem. The agents must run in the cloud because the target devices have little memory or compute to spare. The interaction loop has a hard two-second ceiling, with most behaviour expected much sooner. The system also needs to operate at less than 1% of the inference cost of the small commercial models we benchmarked on the same gameplay workload. Artificial Agency will use its work with Owlchemy Labs on Dimensional Double Shift to explain what makes real-time generative behaviour hard to ship. The talk will cover the gap between cloud reasoning and live game state, the complications created by three concurrent agents, the limits of very small models, and what early playtesting teaches us about making intelligence understandable and fun. The session will include gameplay footage or a live demonstration.
1. Cloud AI for games is a workload problem, not just a per-token pricing problem Real-time games combine bursty traffic, persistent agents, strict response limits, and a steep post-launch demand curve. Infrastructure and pricing products designed for enterprise chat do not map cleanly onto that workload.
2. The reality gap matters more than model latency A small delay in hearing your own voice can disrupt your brain because what you hear no longer matches what you know your body just did. Agents have a similar problem when the game progresses while the cloud model holds an older or incomplete version of reality. Â Traditional game AI can hide latency, but it can also make the situation worse if its actions are not legible to the generative agent. The result can be loops, contradictory decisions, repeated actions, and behaviour that breaks immersion.
3. Small models save cost but they remove room for error Compression and quantization improve cost and speed, but very small models can fall off sharply outside their training coverage. You want to reduce the size as much as possible, tempered by how much unexpected behaviour the game requires it to handle.
4. Three capable agents are not a team Multi-agent behaviour introduces conflicts and feedback loops that do not appear when agents are evaluated individually. Coordination, contention, and social timing are a whole separate problem.
5. Players need to understand the intelligence before they can enjoy it Early playtesting raises unfamiliar design questions. Do players recognize that an agent is intelligent? What do they expect it to know? Do they treat mistakes as a feature (character behaviour) or a bug? Can they cooperate naturally, or do they spend their time trying to break it?
