agent

记录炼云小说工厂 v0.1 的开源过程，以及我对 AI 长篇写作、agent 工作流和可观察生产系统的一些想法。

When one agent becomes many agents, the bottleneck moves from model capability to task, context, permission, state, and acceptance management.

Context rot is not one disease. Long-context distraction and poisoned-context inertia need different product and engineering responses.

My working model for context engineering: the industry moved from adding more context, to pricing every token, to letting agents decide what should enter the working set.

Adding an agent to a product is not the same as making an agent-first product. The hard parts are permissions, tool boundaries, context, task contracts, and treating AI as a real user.

A product note on consumer AI agents: the better the agent works, the more compute it burns, while model providers keep absorbing the visible feature surface.

A practical framework for agent memory: decide who the memory is for, what should be remembered, how long it should last, and how it will be retrieved before picking a tool.

A skeptical note on Hermes-style agent evolution: useful self-improvement is less about magic autonomy and more about captured feedback, skills, and offline evaluation loops.

Most AI agents do not remember you. They reconstruct enough history inside a context window, then continue as if that were memory.

A short note on why agent skills are useful: they keep capabilities out of the main prompt until the agent actually needs them.

一次实时语音语言学习 agent 的产品复盘，记录 YouTube 素材、Gemini Live API、hints、replan、记忆和 agentic 对话体验的取舍。