Most AI Context Layers Are an Inch Off. That's a Mile Wrong.

This post argues that current AI context layers are insufficient because they primarily focus on design-time context (schema, metric definitions) and are often built downstream on data replicas. The key technical contribution is the emphasis on the critical importance of *runtime context* – the live, changing state of the business created by recent transactions and events. The post highlights failure modes of downstream context layers, such as replication lag, agents not seeing their own writes, inconsistent states from CDC, and disagreements between engines. It proposes that the 'customs house' for AI context belongs where the writes happen – within the operational database itself – to ensure agents reason over the same current data that the business is actively writing to. This enables a single engine and a single version of the truth, reducing the need for cross-system data shuttling and improving the correctness of AI-driven actions, especially when agents are performing write operations.