博客
Cyril 是怎么造出来的
一些工程笔记：为什么 AI 原生的平台必须从地基起就这样设计，而不是加在一套已经存在的系统上。
2026年5月8日 · 阅读时长 6 分钟
Why AI-native platforms win — and what that actually means
Bolted-on AI looks impressive in demos and falls apart in workflows. The reason is structural. Here's the engineering decision that separates AI-native from AI-added-on.
2026年5月20日 · 阅读时长 5 分钟
Every record explains itself: the ai_context serializer
Cyril gives every entity an ai_context serializer — a deterministic, schema-aware view built for the model, not the database. It's the unglamorous core of what makes AI grounding actually work.
2026年6月2日 · 阅读时长 6 分钟
Letting AI act safely needs one audit log, not five
Generative AI answers questions; agentic AI takes actions — and actions are where trust is lost. What makes them safe in Cyril isn't a guardrail prompt: it's one permission model, one audit log, one rollback path.