On-Demand Apps

On-demand apps are the source’s “现炒 App” idea: software capabilities are assembled, generated, or combined when the user needs them rather than fully prewritten as fixed application features. In Vol. 161 从开发自己的 OpenClaw 聊起, Justin Yan uses his personal Open Claw-like agent to describe how fewer than thirty AI Skills can already combine in surprising ways.

Vol. 164 从苹果聊到软件未来：Agentic Software 真的要来了？ adds the platform-review boundary. If agents or vibe-coding tools generate calculators, games, or temporary utilities after installation, a marketplace such as App Store has to decide whether it is reviewing a fixed app or an engine for making many short-lived apps.

Vol. 170 Fable 5 重出江湖，GPT 仍需努力 extends the idea into Token-Driven Software. The hosts imagine interfaces, map views, camera effects, NPC behavior, and software flows being generated from the user’s current context rather than assembled only from prewritten features. This broadens on-demand apps from tool composition into dynamic product experience.

Vol. 165 做客声东击西：「龙虾」和 vibe coding 正如何改变我们的思维 adds a practical user-demand boundary. 王俊玉 imagines simple base interfaces that adapt to different user habits, while Justin Yan and 徐涛 note that most people may not want to continuously customize or improve software themselves. The concept therefore depends on making customization feel like the default experience, not a chore.

Key Claims

• Agent Native Software can make app behavior more dynamic because the agent can choose tools and skills based on the user’s immediate goal.
• The concept depends on Agent Harness quality: the agent needs tools, channel access, memory, and permissions before it can assemble useful capabilities.
• It benefits from Vibe Coding because small custom apps and connectors can be created quickly for one person’s workflow.
• Surprise is both value and risk: unexpected combinations can delight users, but they can also create gray-zone behavior or unreviewed actions.
• Platform governance becomes harder when the on-demand app is produced after review rather than submitted as one static bundle.
• The concept pressures traditional SaaS when generic workflows can be generated on demand, while secure, financial, and high-accountability workflows still need strong Agent Permission Boundaries.
• Dynamic generation increases the need for Model Routing Cost Control because always using the strongest model for every generated surface can make the product uneconomical.
• User appetite matters: on-demand apps may work best when they hide customization effort and give people useful defaults rather than asking everyone to become a product designer.

Connections

• Open Claw and Justin Yan — source example and builder context.
• AI Skills, Agent Self-Evolution, and Proactive Agents — mechanisms that let capabilities emerge and improve over time.
• Human-Agent Collaboration — product question of how people ask for, review, and live with generated capabilities.
• AI Inference Cost Structure — economic constraint on always-on or frequently generated app behavior.
• Token-Driven Software, Generated Work Interfaces, and Fable 5 — Vol. 170’s stronger-model extension from app assembly to generated interaction.
• 王俊玉, 徐涛, and Human Judgment Under AI — Vol. 165’s caution that adaptive software still needs defaults, taste, and user willingness.
• Agentic Software, App Store, and AI Communication Ability — Vol. 164’s dynamic-app and review-boundary case.