Kuaikan Comics Builds Livo, an AI-Native Interactive Story World

Summary

Kuaikan Comics, a Chinese comics platform, is developing Livo, an interactive entertainment product designed around AI-native storytelling. The app is expected to launch on app stores soon; its current 1.0 release is internally described as Beta 3.0 after three major iterations. Livo treats each fictional world as an independently running World Agent that manages schedules, events, weather, fortune, and other state changes. A Soul Agent handles user-facing interaction, while each major character and many NPCs operate as separate language-model-driven agents with their own personalities, worldviews, and states. Memory and a causal storyteller are intended to preserve long-term continuity, connect earlier events to later consequences, and allow characters to develop beyond predefined game-like story trees. The system stores basic memories in a vector database and represents longer-term causal relationships in a graph database, with importance weighting and forgetting mechanisms to prevent uncontrolled growth. Its infrastructure also adds model-service integrations, detailed model and agent traces through tools such as Langfuse and LangSmith, and frameworks including LangChain and LangGraph. Kuaikan says the system contains more than 200 nodes, making inference the main cost driver; the team is exploring prompt and token optimization, model-provider selection, and post-training at high-impact nodes. For example, a proposed user mental-state component could model more than 128 features and activate 32, beginning with supervised fine-tuning and potentially adding DPO or GRPO. The team uses baselines, public and internal benchmarks, online traces, sampled user data, adversarial sets, and bad-case feedback to decide whether changes are ready for wider release. Product and engineering work is organized collaboratively, with creators defining a six-layer character model and a harness that converts familiar literary formats into agent context. The company also uses shared AI specifications, architecture reviews, and human approval to coordinate agentic coding across teams.