Explain why context engineering emerged as a distinct discipline in 2025-2026

In 2022 and early 2023, the phrase 'prompt engineering' covered the entire job of getting good behavior out of a language model. The work was almost entirely about how to phrase the instruction. By 2025-2026 the work has split into two clear disciplines, and three concrete drivers caused that split. **Retrieval and tool use became table stakes.** Almost every non-trivial application in 2026 pulls in retrieved chunks from a vector store, runs tool calls, or both. A typical RAG application sends 2K-10K tokens of retrieved content per turn. A typical agent application accumulates tool outputs over many turns. The instruction itself is a small fraction of what reaches the model, and the rest is composed at runtime from databases, search results, tools, memory, and history. Tuning the instruction in isolation cannot fix problems that live in that composition. **Context windows grew, but effective accuracy did not keep up.** In 2022 a context window was a few thousand tokens. In 2026 frontier models advertise 200K to 2M token windows. But the empirical evidence, the lost-in-the-middle finding (Liu et al. 2023), long-context benchmarks like RULER (2024) and BABILong (2024-2025), consistently shows accuracy degrading well before the hard limit. The result is that there are now two budgets, not one: the hard limit the API will accept, and the soft budget beyond which quality starts to fall. Managing the soft budget is a first-class engineering concern that did not exist when the hard limit was 4K. **Agentic loops produce long trajectories with state.** Agents that call multiple tools across many turns produce trajectories whose history, scratchpad, plan, and tool results all need to be selectively kept, summarized, or evicted. This is state management across iterations, not instruction writing. The skills that make a good agent loop work, placement, eviction, summarization, cache-aware design, are not prompt-engineering skills. The common thread across all three drivers is that the work shifted from the words inside the instruction to the system that composes the full input. That is what the new term names.

Three drivers, each with a concrete signature in how production LLM applications are built. **Retrieval and tool use became default architecture.** In 2023, RAG and tool use were emerging patterns; by 2026 they are table stakes for almost any non-trivial application. A modern call to a frontier model typically includes retrieved chunks from a vector store (2K-10K tokens, sometimes more), tool outputs from earlier steps, and persona facts from a memory store. The instruction itself is often a small fraction, sometimes under 5%, of what reaches the model. This shifts the locus of design from *how is the instruction phrased* to *what gets composed into the input on this turn*. The vector store, the rerank pipeline, the tool wrappers, the memory store, the schema definition, none of these are prompt-string problems, and none of them are solvable by tuning the instruction. **Context windows grew but effective accuracy did not.** From 4K in early 2023 to 200K-2M in 2026, the advertised context limit has grown by orders of magnitude. The empirical reality is more constrained. Lost-in-the-middle (Liu et al. 2023) showed that LLM attention is U-shaped over long contexts and content in the middle gets attention-starved. Subsequent long-context evals, RULER, BABILong, NoCha, replicated and extended this finding across every modern frontier model. The implication is that the *effective* context budget is much smaller than the *hard* limit, and the soft budget is task-dependent, model-dependent, and usually a fraction of the advertised number. Managing the soft budget, deciding what goes in, what gets evicted, what gets summarized, where each slot is placed, is a first-class engineering discipline that did not exist when budgets were small enough to use everything that fit. **Agentic loops produce long stateful trajectories.** Agent applications run for tens or hundreds of iterations, with each iteration adding tool calls, observations, scratchpad notes, and plan