Inside a Game Where AI Agents Never Sleep
Isaac Lee of NEXUS explains why he wants to fill MMORPGs with agent players instead of scripted NPCs to bootstrap multiplayer worlds, and walks through NEXUS's AI lab: text-to-3D world building, novel-to-video pipelines, pixel animation loops, the Da Vinci multi-agent dev framework, self-hosted inference after Kimi, the Bob meeting assistant, and the implications of agentic shopping for e-commerce.
Agent Populations in MMORPGs and the Road Beyond Studio and User-Generated Content
Isaac Lee, who runs AI and blockchain experiments at NEXUS, opens by framing the core bet: populate an MMORPG with agent players instead of scripted NPCs, so a game feels alive from day one. He contrasts this with traditional NPCs, where “everything’s scripted, right? So there’s no spontaneity.” His point is that an agent’s interaction with you is shaped by time, place, and its own reasoning, so it cannot be replayed the way a dialogue tree can.
The Evolution of Game Content
Isaac lays out a three-stage history of game content. Studio-generated content was bounded by what developers defined, and “once you finished the contents that was it.” User-generated content came next, giving players an open world to build maps, features, and loops. He argues the next phase is agent-generated content, where agent populations make the world “vast and dynamic” as you move through it.
Solving the Chicken-and-Egg Problem
A key motivation is bootstrapping. Multiplayer launches resemble marketplaces: you need players to attract players. If agents fill the world on launch day, the tipping point is already passed, and human players arrive into a world with competition, social dynamics, and spending pressure. Isaac links this to monetization and game “half-life,” noting that successful MMORPGs retain users because of competition to spend and get ahead. Agents could extend that lifetime value as the game matures and inflation sets in.
On design constraints, Isaac suggests agents should start from the same ground as players, with the same restrictions, starting gear, quests, and side quests, and the freedom to choose which to pursue.
Relentless Agents and Auto-Play
When asked about the fact that agents never sleep, Isaac points out that MMORPG players already rely on auto-farming and auto-mode. Players cherry-pick the moments they want, “the dopamine loop,” while the character runs on autopilot otherwise. From that angle, agents extend an existing pattern rather than introducing something alien, and they might lengthen the experience and add variety.
AI Companions and Coaching
The conversation starts with a GDC presentation from a Tencent subsidiary showing metrics for in-game AI playmates that learn your play style and act as companions, boosting retention. Isaac notes that Krafton, operator of Battlegrounds, is researching this too. The idea is less a rival character and more an omnipresent coach that identifies where you tend to die based on your history. Its form depends on genre: an FPS companion coaches tactics, while an MMORPG benefits from populated agent players.
The Last 10 Percent
A recurring theme is the gap between AI-generated prototypes and shippable products. Isaac says AI gets a product to roughly 80 to 90 percent, but the final stretch is where NEXUS focuses effort, because “the users doesn’t really care whether you use AI.” The host connects this to the graveyard of unfinished projects. Isaac’s examples include:
- Text-to-3D world builder: A tool that gives AI a preset of engines and integrates existing physics engines (including Roblox’s) so creators can prompt a town or map that is actually playable. The hard part is defining in advance what the model should do, such as what a city must contain and what capabilities a character should have, which Isaac describes as “very similar to a tool call.”
- Novel-to-web-drama pipeline: Through its acquisition of ONE store, NEXUS has access to comics and novels. Models currently generate roughly 20 to 30 seconds of usable video, so a series needs hundreds or thousands of clips that still read as a coherent story. Clips generated from similar prompts vary, so the team writes highly specific instructions covering scenario, character appearance, wardrobe, and dialogue behavior, then iterates heavily.
- Pixel animation loop: Generating pixel art is easy, but natural motion is hard. The first attempts were “awful.” The fix was to specify what the character should do and run an agent loop that keeps filling the gap until the motion reads naturally.
The discussion also touches on fuzzy language. Sliders and histograms give fine granularity that “make it more pink” does not, so specification becomes the real work.
Da Vinci and Agentic Development
NEXUS built Da Vinci, an internal multi-agent development framework, which the team uses to ship products. The workflow classifies intent (for example, “build a note-taker app”) against a blueprint, then dispatches role-based agents for backend, frontend, feature definition, and QA. A loop-based orchestrator reviews outputs, identifies missing pieces, and iterates. Isaac says much of the human time now goes into designing core architecture and evaluating output, while deployment is fully automated. Before pushing, he has an agent click every button and link, produce a screenshot report, and spawn sub-agents to fix broken elements.
The QA discussion leads to a idea: undercover QA agents playing inside the game world, reporting bugs and triggering fix loops. NEXUS has not tried this yet. Isaac still believes human judgment is essential, because the goal is user experience and the “magic” that hooks a player sits beyond the polished 90 percent. He summarizes: “Games are games are for humans, I would say for now.”
Custom Harnesses Versus Frontier Tools
Isaac explains why NEXUS built its own framework: it can hand-pick features, the design of the process can be tuned to its own needs, and security teams prefer closed systems. He defines a harness as anything from a simple instruction such as “when prompted A, look at skill B, then do C in order” to a complex structure. NEXUS’s first harness defined around 200 agent roles. As models improved, that scaffolding became unnecessary. Harnesses still help when a team knows exactly what it wants and needs repeatable quality, but they are getting thinner. The host notes the industry moving toward minimal harnesses, where you give a model tools like bash, GitHub access, or Claude Code and Codex, and let it decide. Isaac agrees that Claude Code and Codex are themselves good harnesses, and that custom ones matter when a specific need arises.
Frontier Models, Kimi, and Self-Hosting
NEXUS currently develops mostly with frontier-lab models, which Isaac believes remain somewhat better than open-source. The Kimi release changed the calculation. If the team can reach 80 to 90 percent of its work at a tenth of the token cost, it is worth exploring, so NEXUS has started building a small internal data center. Isaac is focused on inference rather than training, and he estimates GPUs could last around five years for that purpose. The host adds that engineering time to operate GPUs must be counted, but that depreciation over a long horizon favors owning hardware, especially since models can be swapped as they improve. Older hardware such as A100s remains usable, and model routers can direct smaller models to older machines.
Isaac raises a valid concern: open-source labs may stop releasing weights, as Meta did before reversing course. He still thinks the experiment is worth running.
The discussion turns to the difficulty of the infrastructure itself. The host’s Swiss friends found GPUs “very finicky” and needed low-level expertise, since NVIDIA is not forthcoming about what sits beneath CUDA. They recommend a book on inference optimization written by someone who reverse-engineered GPU behavior, along with vLLM for faster serving, and communities such as GPU Mode. They also cover neoclouds: Modal, which offers fast serverless GPU spin-up, and Baseten, which has built an optimized inference stack on top of vLLM.
Bob, Personalization, and Context
NEXUS built Bob, a meeting note-taker running on a local GPU. It captures discussion, flags technical claims such as GPU costs or inference expenses, and verifies important facts in a side chat through another model call. Bob is still under development and expected to ship internally within weeks.
Isaac wants Bob to know who each person is: their role, tendencies, and preferred note format. Sales and engineering would receive different styles. The host suggests building profiles from discussions, embedding chunks in a vector store, or something simpler like a file system with per-person skill files that describe how someone likes to receive information.
The conversation then turns to how much context is too much. The host is willing to feed an AI all of his data, including Apple Watch and health-band readings, for daily reports. Isaac notes that workplace context raises different concerns, such as an employer seeing stress patterns during standups. They discuss a woman who analyzed her dating history with Claude and faced backlash, and they agree the boundary is unclear. Isaac points to a growing presence of AI observers, comparing it to how ubiquitous phone cameras have normalized being photographed. The likely result is less privacy and more AI in daily life. Whether that is dystopian or useful “depends on how you use it.”
Marketing Agents and the Agentic Shopper
NEXUS also built a marketing agent that monitors social mentions, mainly on X, analyzes positioning, and suggests actions in a loop. It has analyst agents and data-collection agents. The team has used computer-use approaches to gather social listening data where APIs are restricted, such as LinkedIn, though this is slower and carries risk if platforms respond. Isaac and the host see that blocking agents may cost sites traffic, which makes the trade-off a double-edged sword.
Shopping is the early frontier. An agent with access to health data, preferences, and history might recommend eye drops directly by API. The host recalls Amazon suing Perplexity over a computer-use shopping agent, and a court ruling that allowed Perplexity to operate such an agent, which weakens Amazon’s position and its first-party data relationship. The deeper issue is design. E-commerce sites are hyper-optimized for human psychology, with popups, email capture, and free-shipping thresholds. An agent that simply wants pants ignores all of that. Isaac suggests providing straightforward information and identifying who the calling agent represents, perhaps through authenticated access. Sponsored content and ad revenue become open questions as agents mediate purchases.