Context failure occurs when an AI system produces the wrong result because its supplied information is deficient, mismatched, stale, distracting, or poorly positioned. It is the most directly controllable failure category in agent-assisted work, and it follows from treating an AI model as an oracle instead of a programmable text engine. Context engineering addresses the whole information environment: what the model receives, in what order, at what detail, and from which sources. section 3.0 section 3.1
Recognize context failure in four recurring forms:
- Too little context creates a vacuum in which the model must guess missing requirements, dependencies, or intent. section 3.0
- Too much context creates context rot, where irrelevant information competes with the task. section 3.0
- Stale or incorrect context poisons the result, such as directing an agent toward obsolete documentation. section 3.0 section 3.5.4
- Correct information can still fail when it appears where the model has stopped paying attention: lost in the middle. section 3.0
Treat context as an engineering input rather than a prompt-writing detail. Set the goal, constraints, and acceptance criteria; supervise retrieval; correct wrong sources; and maintain persistent role and rule context through an agent manifest. Agents reduce manual fetching, but they do not remove the need to direct, inspect, and debug the context-building process. section 3.5.4 section 3.0
