Technology Hub
How Kult Built a Game Companion That Understands the Player’s Context
In the middle of a game, getting help often means breaking the moment: pause, alt-tab, open a browser, search a guide, and hope the answer matches the current patch and the player’s progress. Google…
In the middle of a game, getting help often means breaking the moment: pause, alt-tab, open a browser, search a guide, and hope the answer matches the current patch and the player’s progress. Google Cloud’s latest customer case study describes a different approach: an assistant that sits beside the game, listens to the player’s question, understands the current context, and replies through voice and a second screen.
On 20 November 2026, Google Cloud published a Blueprint case study about Kult, a London-based startup building this kind of gaming companion with Gemini models, a structured knowledge base, Google Kubernetes Engine, and a native visual feed. The interesting technology lesson is not simply “add a chatbot to a game.” It is how the team treats latency, accuracy, cost, context, and interface control as one product problem.
The problem is more specific than game chat
According to the Google Cloud case study, Kult wants players to ask questions aloud while keeping their hands on the keyboard or controller. The answer can arrive in voice, while a visual response appears on a second screen with elements such as images, video, maps, builds, interactive tools, and links.
That design changes the product requirement. A general-purpose assistant can answer a broad question, but a useful game companion has to know which game is running, where the player is, what they have been doing, and what they asked a moment ago. “What is he weak to?” only makes sense when the system knows which boss the player is facing.
The answer also needs to reflect the game’s current state. A patch can change items, abilities, routes, or enemy behavior. A response based on a general model’s memory may sound confident while being outdated. Kult’s approach, as described by Google Cloud, is to build and maintain a structured knowledge base for each covered game and have the model reason over that knowledge.
Why latency, accuracy, and cost pull in different directions
Kult identifies three constraints. Latency matters because a slow reply feels disconnected during play. Accuracy matters because the wrong build or strategy wastes time and may send a player in the wrong direction. Cost matters because an always-on companion may handle many turns in one session, and a consumer product cannot make every turn expensive.
These constraints make a single-model architecture unattractive. A large model may be valuable for deep reasoning, but using it for every classification, lookup, and conversational turn can increase both cost and response time. Kult’s solution is a model range. Google Cloud says a lighter Gemini Flash tier handles high-volume, repetitive work, while more advanced models handle rarer, heavier reasoning performed for each game.
The principle is portable beyond gaming: use the smallest model that can reliably complete the step, and reserve larger models for decisions where the extra reasoning changes the result. That requires routing, evaluation, and an escape path when the smaller model is uncertain.
Routing is a product feature
Kult’s backend uses a routing layer to select a model for each task, with the service running on Google Kubernetes Engine. Google Cloud’s case study says the team uses a decision-only model called Jev for choices such as whether a request needs a lookup or whether an answer covered the point. Confident cases can settle quickly; uncertain cases escalate to a larger Gemini model.
For developers building a similar system, the important idea is to make routing observable. Log which model handled each step, why the router chose it, how confident it was, and whether a human-authored reference answer would have preferred another path. When a model is replaced, compare it against a fixed evaluation set before changing production traffic.
That turns model selection from a hidden implementation detail into an engineering decision. It also helps teams explain cost changes. A rise in spending might come from more user sessions, more escalation to the large model, a larger context window, or a knowledge-base retrieval problem that creates repeated calls.
The interface is constrained on purpose
A voice answer is useful while the player is focused on the game, but voice cannot carry every detail. Kult’s second-screen feed can show richer material, yet the model is not allowed to write arbitrary front-end code. Instead, Google Cloud says the model describes the interface it wants using a catalog of components designed by the team, and the application renders those components in Kult’s own style.
This is a strong pattern for agent-generated interfaces. The model can choose from approved building blocks—such as a map, item list, guide card, or interactive tool—while the application controls layout, behavior, and safety. The result is less flexible than allowing generated code, but it is easier to test, version, and keep consistent across devices.
For a small product team, a component catalog also creates a boundary between experimentation and the user experience. New visual responses can be added deliberately, without giving every model response the power to change the application’s interface.
Keeping context fresh is the hard part
Google Cloud says Kult now covers more than 100 games and that adding a game is an automated process rather than a manual project. That claim points to the real scaling challenge: knowledge maintenance. Each game can have its own vocabulary, patch cycle, data sources, and community expectations.
A repeatable ingestion and evaluation process matters more than a large initial list of games. New content should be checked for freshness, structured into a searchable representation, and tested against reference questions written by people who know the game. When a patch changes an answer, the system needs a way to identify affected knowledge rather than silently serving yesterday’s guidance.
The team also describes a feedback loop in which player questions reveal gaps in the knowledge base. That can guide what to improve next, but it should be treated as product telemetry with clear privacy and retention boundaries. What players ask can be valuable; it can also be personal.
A practical blueprint for similar products
- Define the user’s immediate context before choosing a model or building a prompt.
- Separate fast conversational turns from slower research or reasoning tasks.
- Use a structured, versioned knowledge source instead of relying on model memory alone.
- Render agent-generated interfaces from an approved component catalog.
- Evaluate routing decisions, answer freshness, latency, and cost together.
- Keep a clear path for user feedback, correction, and privacy controls.
Kult’s architecture is a useful case study because it treats an AI companion as a system, not a model call. The voice response, second-screen UI, knowledge base, model router, deployment platform, and player context all contribute to whether the product feels helpful. The broader lesson is simple: when an assistant lives inside a time-sensitive workflow, usefulness comes from fitting the moment—quickly, accurately, and within a cost structure the product can sustain.
This article is a general information summary based on the cited public sources. It is not legal, cybersecurity, or compliance advice.