Token-Driven Software

Token-driven software is the episode’s speculative frame for software whose behavior, interface, content, or world rules are generated at use time by models rather than fully fixed ahead of time. In Vol. 170 Fable 5 重出江湖，GPT 仍需努力, Justin Yan and 自立 imagine apps adapting to the user, weather, device state, lighting, and momentary context; maps that generate new visual layers as users click; camera or AR effects that produce unexpected but reasonable feedback; and games whose NPCs, combinations, and sandbox rules can be generated instead of exhaustively prewritten.

This extends On-Demand Apps and Generated Work Interfaces from work tools into a broader product form. The core tradeoff is that token-driven experiences can be more personal, surprising, and exploratory, but they inherit AI Inference Cost Structure, latency, quality-control, safety, and product-taste problems.

Vol. 164 从苹果聊到软件未来：Agentic Software 真的要来了？ adds the prelude: before the hosts name token-driven software explicitly in Vol. 170, they describe Agentic Software as a move from fixed product paths toward personal, scene-specific software assembled from capabilities and context.

Key Claims

• The shift is from fixed screens and predefined branches toward model-generated interaction surfaces and local context-sensitive behavior.
• AI-native operating systems may be a more natural carrier than ordinary app wrappers if the system needs device state, permissions, and continuous context.
• Games and interactive entertainment are natural testbeds because users already value exploration, surprise, NPC behavior, and world-rule experimentation.
• Research and simulation may also benefit when generated virtual environments reduce the cost of testing ideas before physical experiments.
• The product challenge is not only generation; the experience must still be coherent, reliable, emotionally rewarding, and economically sustainable.
• Token-driven software can make copying surface UI easier, pushing defensibility toward workflow depth, data, taste, distribution, and proprietary execution.
• Earlier agentic-software speculation points in the same direction: dynamic software needs capability atoms, context, and review surfaces, not only generated UI.

Connections

• On-Demand Apps and Agent Native Software — adjacent software forms where agents assemble capabilities at need time.
• Generated Work Interfaces and Agent-Facing Interfaces — interface patterns that make generated views and callable actions useful.
• AI Interactive Entertainment, AI Interactive Content Platforms, and Designed Agency In Games — game and entertainment layer.
• World Models, Causal World Models, and AI For Science — generated-world and simulation implications.
• AI Inference Cost Structure, AI Subscription Economics, and Model Routing Cost Control — cost constraints on dynamic generation.
• Agentic Software, Atomic Capability Services, and AI Communication Ability — Vol. 164 foundations for dynamic software behavior.