Context engineering—the practice of optimizing data fed to AI agents—is necessary but insufficient for production agentic systems. AI agents operate in observe-decide-act loops that require stateful, continuously updated business objects (e.g., Customer, Order, Inventory) derived from multiple operational sources. Traditional infrastructure options (operational databases, data warehouses, stream processors) each fail at least one of the three non-negotiable requirements: freshness, correctness, and composability. A 'context engine' is proposed as the missing infrastructure layer: a system that pre-computes and continuously maintains live business data objects queryable in milliseconds via MCP. Materialize is presented as an implementation of this pattern, using incremental view maintenance and SQL to serve always-fresh, always-correct context to AI agents.
Table of contents
Agentic context is statefulWhat is a context engine?Context engineering vs the context engineA context engine runs on a live data layerBuilding the context agents actually needFire up your context engineSort: