context engineering

Rod Johnson called Context Engineering "a valuable advance over prompt engineering" on 23 July 2025, then argued the definition was incomplete: filling the context window with relevant information says nothing about what comes back out of the model, nor about the business systems an LLM application must plug into. His extension, DICE, structures both inputs and outputs with a domain model.

2025 is the dominant period, and the term has been under pressure since. On BFM Business in May 2026, Rémi Jacquet laid out a pivot: prompt engineering in 2024, Context Engineering in 2025, harness engineering in 2026. Lushbinary's guide stacks the same sequence instead of retiring its steps: prompt, then context, then Loop Engineering, each layer containing the one before, the complexity unchanged and only the leverage moving.

The Google whitepaper by Addy Osmani, Shubham Saboo and Sokratis Kartakis names six context types (instructions, knowledge, memory, examples, tools, guardrails), split between always-loaded static context and on-demand dynamic context. Hugo Lassiège stratifies his by loading moment: a short permanent CLAUDE.md, rules activated by path, personas opened only when needed. Assaf Elovic measured what filtering is worth on the retrieval side: his Jev filter sends 4,500 tokens instead of 7,600, for an equivalent quantity of relevant text.

Matthieu Grymonprez supplies the limit. "Un agent n'a pas de bon sens": a mediocre definition gets applied exactly as written, including where it holds for France and not for Spain.