The successor to prompt engineering. Once systems started retrieving documents, calling tools and carrying memory, the interesting question stopped being how to phrase a request and became what should be in the context window at all, in what order, and what should be dropped. Most quality problems in serious applications turn out to be context problems rather than model problems.
Technical view
Concerns include budget allocation across system instructions, memory, retrieved material, tool definitions and conversation history; ordering, given position effects such as lost in the middle and the value of a stable cacheable prefix; compaction and summarisation as a conversation grows; and pruning of stale tool results. Tool definitions are themselves part of the context and degrade selection accuracy when too numerous, so exposing a smaller relevant set matters. The discipline is closer to systems design than to writing, and it is where most engineering effort in agentic applications now goes.
Related
Part of AI Terminology Index, section: Controlling and Extending Behaviour at Run Time.