Klang Builds SEED With Persistent AI NPCs
80 Level reports that Klang Games' society simulator SEED runs thousands of AI-controlled characters that keep working, building relationships, and participating in its virtual society when players are offline. The August 5 report describes a custom inference pipeline behind the system, while Klang's earlier technical overview identified free-form player chat and evolving character relationships as generative AI use cases.
Klang Games' SEED uses thousands of AI-controlled characters to maintain a persistent virtual society around the clock, according to an August 5 report by 80 Level. The report describes a game in which players manage AI-driven characters called Seedlings, whose simulated lives include careers, relationships, families, and contributions to player-built cities even while their owners are offline.
80 Level characterizes the system as a society simulator inspired by the social ecosystems of EVE Online. Mundi Vondi, Klang Games' CEO, told the publication that the team sought to make large-scale online social systems more approachable by making the characters in the world persist beyond player sessions. "If every character continued living around the clock, then suddenly player-built cities wouldn't become ghost towns whenever people logged off," Vondi said.
AI beyond dialogue
The reported system goes beyond chatbot-style conversation. According to 80 Level, Seedlings contextually understand their surroundings and can be spoken to while they carry out their simulated lives; the publication says Klang uses a custom-built inference pipeline to support those behaviors.
Klang's May 2025 product post provided further detail on the features it was developing. The company wrote that its Game AI system shapes Seedling personalities, motivations, and relationships over time, and that generative AI is combined with player choices to produce evolving stories. Klang also described a Seedling-to-player chat system for free-form conversations, including interactions in which players can check a character's well-being and offer guidance.
Neither the 80 Level report nor Klang's May 2025 post specifies the models, inference hardware, latency targets, evaluation methods, or guardrails used by the pipeline. Those omissions leave open practical questions about how the game constrains generated interactions, maintains behavioral consistency across long-running agents, and controls operating costs at the reported scale.
Persistent agents raise systems questions
For ML and game-engineering teams, SEED is an example of a broader design pattern: using generative systems alongside stateful simulation rather than treating an LLM as an isolated dialogue layer. In comparable persistent-agent systems, reliable world state, memory management, action validation, and cost-aware inference routing are as consequential as language quality.
The player-facing risk is also visible in contemporaneous coverage. PC Gamer's article focused on immersion failures caused by weak chatbot dialogue. That critique is not a technical evaluation of Klang's underlying pipeline, but it underscores a common production constraint for AI characters: free-form interaction needs to remain coherent with game state, character identity, and player expectations over repeated sessions.
In persistent-agent systems, engineering emphasis often falls on continuity. A world containing many autonomous characters requires generated language and decisions to fit within a durable simulation, rather than merely producing plausible responses one prompt at a time.
Key Points
- 180 Level reports that SEED keeps thousands of AI-controlled Seedlings active after players leave, extending simulation beyond conventional session-bound NPC behavior.
- 2Klang's 2025 overview links generative AI to free-form chat, evolving relationships, and player-influenced stories, not dialogue generation alone.
- 3Comparable persistent-agent systems typically require durable state, action validation, memory controls, and cost-aware inference alongside convincing generated dialogue.
Scoring Rationale
SEED is a notable applied-AI experiment in persistent game-world simulation, with relevance for teams building stateful agent experiences. The available reporting does not disclose models, benchmarks, costs, or deployment architecture, limiting its immediate technical reproducibility.
Sources
Primary source and supporting public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
