# Inside a Game Where AI Agents Never Sleep

Issac Lee, NEXUS | MLOps Community | 1:01:04
Hosted by Demetrios Brinkmann

Source: https://www.youtube.com/watch?v=H8MNuoxAmNU
Channel: MLOps Community, now AAIF Live (https://www.youtube.com/@AAIFLive-x1r). Summarised by MLOps Talks.
Page: https://mlopstalks.com/talks/inside-a-game-where-ai-agents-never-sleep
Published: 2026-10-08
Tags: agents, inference, multi-agent, privacy

## TL;DR
- NEXUS is testing agent players in MMORPGs to create social activity before a game has enough human players.
- The team has reduced its internal agent framework from about 200 defined roles as models have become better at deciding how to work.
- NEXUS is considering a small internal data center to run open models for inference after the Kimi release changed the cost calculation.

## Summary
Issac Lee describes NEXUS experiments with AI-generated games, agent players, development tools, and personal context. The team is building text-to-3D tools for Roblox and systems that turn existing novels into short video dramas. Lee argues that AI can produce much of a game or media product, but the last part still needs detailed instructions, repeated evaluation, and human judgment. NEXUS is also testing agent players in MMORPGs. These agents could create social activity at launch, behave less predictably than scripted NPCs, and keep performing tasks while human players are offline. For software development, NEXUS uses Da Vinci, an internal framework with agents for backend work, frontend work, feature definition, and QA. Lee says its large set of hand-defined roles has become less necessary as models improve. The discussion ends with local inference, meeting notes that remember user preferences, privacy, marketing agents, and shopping agents that may force ecommerce sites to support machine customers.

## Key ideas
### AI companions can adapt to how a player already plays
[01:17](https://www.youtube.com/watch?v=H8MNuoxAmNU&t=77s)
Issac Lee describes AI companions that have access to a player's game history. A companion can act as a playmate, help with decisions, or coach the player based on repeated weaknesses. He cites research by Krafton, the operator of Battlegrounds, where an agent can tell a player that they are vulnerable in a certain situation and connect that advice to past play. The companion does not need to operate at maximum difficulty. It can meet the player at an appropriate level and respond to commands such as waiting in a particular place.

### Agent players could give an MMORPG a social population at launch
[06:27](https://www.youtube.com/watch?v=H8MNuoxAmNU&t=387s)
NEXUS is studying MMORPGs where some of the characters are agent players rather than scripted NPCs. Lee says a successful MMO depends on competition and interaction among players, but a new game starts with too few people to create those dynamics. Agents could fill the world from day one and help solve that chicken-and-egg problem. They would have the same starting restrictions as human players and could take part in quests and other activities. Lee expects this could also extend the useful life of a game as its population and economy change.

### Agent-generated content follows studio-made and user-generated content
[08:24](https://www.youtube.com/watch?v=H8MNuoxAmNU&t=504s)
Lee describes three stages of game content. In the first, a studio defines everything in a fixed setting. In the second, users receive an open world and create maps, features, and gameplay loops. The next stage, in his view, is agent-generated content inside the game. Agent players could make interactions less predictable than scripted NPCs because the response could depend on the character, the situation, and the history of the interaction. This could create a larger range of experiences without requiring a studio to define every outcome in advance.

### The last part of an AI-generated product needs precise instructions
[14:43](https://www.youtube.com/watch?v=H8MNuoxAmNU&t=883s)
Lee says AI often gets a product most of the way toward completion, but the remaining work determines whether people can actually use it. For a novel-to-video project, NEXUS may need hundreds or thousands of short clips that must look like one continuous story. Similar prompts can produce inconsistent characters and scenes, so the team specifies the order of operations, character appearance, clothing, direction of gaze, and sound. NEXUS applies the same approach to pixel animation. It defines the desired motion and uses a loop in which an agent repeatedly works on the missing details.

### Da Vinci coordinates specialized agents and repeated reviews
[19:36](https://www.youtube.com/watch?v=H8MNuoxAmNU&t=1176s)
NEXUS uses an internal development framework called Da Vinci. When an engineer asks it to build an application, the system classifies the task and applies a blueprint. Backend agents, frontend agents, feature agents, and QA agents receive different jobs. A coordinating process assigns work, reviews the result, identifies missing pieces, and starts another iteration. Lee says deployment has become largely automated. Before pushing an application, he asks an agent to click buttons and links, take screenshots, and write a report. If it finds a broken interaction, another agent can try to fix it.

### Better models have made large hand-defined harnesses less necessary
[28:12](https://www.youtube.com/watch?v=H8MNuoxAmNU&t=1692s)
Lee says NEXUS once defined about 200 agent roles and detailed instructions for how each one should work. As models improved, many of those definitions became unnecessary. A harness can be small, such as telling an agent to consult a particular skill and follow a particular order, or it can define a complete repetitive process. Detailed harnesses still help when a team needs consistent outputs or has a very specific workflow. Lee expects their importance to decrease as general models become better, while custom harnesses remain useful for specialized needs.

### Local inference may make economic sense for a high-volume development team
[32:12](https://www.youtube.com/watch?v=H8MNuoxAmNU&t=1932s)
NEXUS mainly uses frontier models, but the Kimi release led the company to consider running models itself. Lee says NEXUS spends a lot of tokens each month and is evaluating whether it could achieve most of its current work at a much lower cost with local inference. The team has started looking at a small internal data center with GPUs. It is focused on inference rather than training, and Lee estimates the hardware could remain useful for years. He also names open-model availability as a risk, since model providers can change whether their weights remain open.

### Personal context makes AI more useful and raises uncomfortable privacy questions
[46:36](https://www.youtube.com/watch?v=H8MNuoxAmNU&t=2796s)
NEXUS is developing Bob, a meeting note taker that runs on a local model and can ask another model to verify important points during a discussion. Lee and Demetrios Brinkmann discuss giving Bob profiles of recurring participants so it can format notes in each person's preferred style and remember their work context. They also question how much personal information an AI should receive, including health data, meeting behavior, and records of private conversations. Lee does not claim to have a clear boundary. He says people may have to get used to AI observers that study interactions and suggest changes.

### Shopping agents may change what ecommerce sites are designed to do
[56:57](https://www.youtube.com/watch?v=H8MNuoxAmNU&t=3417s)
Lee expects shopping to be an early widespread use of agentic systems because an agent can combine information about a person's food, health, habits, and current needs. NEXUS is also building a marketing agent that monitors discussions about a product on X and suggests how to position it. Brinkmann argues that ecommerce pages are designed around human psychology, with email prompts, upsells, and free-shipping thresholds. A shopping agent may ignore those tactics and go directly to the requested product. Lee says sites may need to decide whether they are optimizing for human users, agents, or both, and may need to identify the agent making a request.

## Notable quotes
- Issac Lee: "If you populate an MMORPG with agent players instead of NPCs, you're able to have a much more vast and dynamic experience." (00:00)
- Issac Lee: "The thing is, they're relentless. A player has to sleep at some point. An agent doesn't." (11:10)
- Issac Lee: "As the models got better and better, this becomes a lot less necessary." (29:29)
- Issac Lee: "Games are for humans, I would say for now." (24:42)
- Issac Lee: "I think one of the earliest adoption of the agentic future would be shopping." (56:57)

## Tools & references mentioned
- NEXUS
- MLOps Community
- Demetrios Brinkmann
- Roblox
- Krafton
- Battlegrounds
- Da Vinci
- One Store
- Kimi
- Claude
- Codex
- ChatGPT
- vLLM
- NVIDIA
- A100
- H100
- X
- Perplexity
- Amazon
- CUDA

## Who should watch
- You are building AI features for games and need examples of companions, agent players, or AI-generated worlds.
- Your team is deciding how much orchestration and role definition to keep as foundation models improve.
- You are weighing local inference, personal context, or machine-driven shopping against the cost and privacy tradeoffs.

## Related talks

- [AI Agent Development Tradeoffs You NEED to Know](https://mlopstalks.com/talks/ai-agent-development-tradeoffs-you-need-to-know) (Sherwood Callaway, 11X, 57:07)
- [Building Multi-Player AI Systems (and why it's SO hard)](https://mlopstalks.com/talks/building-multi-player-ai-systems-and-why-its-so-hard) (Tula Masterman, MeshAgent, 13:53)
- [Real World AI Agent Stories](https://mlopstalks.com/talks/real-world-ai-agent-stories) (Zach Wallace, Nearpod Inc., 47:08)
- [The Agent Landscape - Lessons Learned Putting Agents Into Production](https://mlopstalks.com/talks/the-agent-landscape-lessons-learned-putting-agents-into-production) (Paul van der Boor & Floris Fok, Prosus Group, 1:08:41)
- [A New Way of Building with AI](https://mlopstalks.com/talks/a-new-way-of-building-with-ai) (Jiquan Ngiam, Lutra AI, 1:04:46)