Context Engineering

Prompt engineering focuses on instructions and examples. Context engineering also decides which records, history, tool definitions, and tool results enter the active request. Memory addresses what persists and how it returns later. These responsibilities overlap in an agent application.

Example: for a support question, provide the applicable policy and order details that the user is authorized to access. Supplying the entire customer database adds irrelevant information and creates an access problem. Choosing scope is part of the system design.

For long tasks, preserve goals and constraints while managing growing history. In episode 53, at 36:33, Richmond Alake discusses compression in the active window and retrieval from an external store. His terminology differs somewhat from other systems, so the important distinction is what remains in context and what must be fetched.

Context also needs maintenance after launch. In episode 75, at 45:14, WisdomAI CPO Kapil Chhabra calls this the context development lifecycle: gather the context an organization has documented, make it usable by agents, and watch for business definitions that change. In his earlier example, the system surfaces evidence that a shared definition may be out of date, and a subject-matter expert decides whether to revise it. WisdomAI sponsored that interview, and the lifecycle is Chhabra’s name for this maintenance work rather than an industry standard.

Conor’s context essay gives a concrete example of excessive tool output. Measure whether a context change improves task results; more input and a more elaborate architecture are not benefits on their own.

See AI vocabulary in plain English for this term alongside the other language of models, data, agents, and infrastructure.