Awesome Context Engineering

★ 3.3kContext Engineering

#awesome#context#research

Overview

A curated research survey of 'context engineering' (3.3k+ stars) — not a hands-on tutorial, but a knowledge map organizing the field's research papers, blogs, tools and practices by topic. The authors also published a companion survey, 'A Survey of Context Engineering for Large Language Models' (arXiv 2507.13334, covering 1400+ related works), and this README serves as the browsable index to it. A great starting point for systematically learning the field or finding an entry point in any direction.

Who it's for

Researchers and engineers who want a systematic view of the context engineering landscape, including those tracking the 2026 shift from context engineering toward agent engineering.

What you'll learn

01What it is: a survey list + a survey paper, not a tutorial
It follows the Awesome-list convention but is more academic than most: each category lists papers, benchmarks, tools and blogs with arXiv or venue sources. It pairs with a published survey, 'A Survey of Context Engineering for Large Language Models', which systematically analyzes 1400+ related works; the repo's structure mirrors the paper's sections — reading the README is reading a clickable, deep-diveable survey.
02What it collects: resources organized by topic
Content is filed by topic: context scaling (position interpolation and ultra-long context), context management in production (compaction, caching, scoped instruction loading), structured data integration and knowledge graphs, self-generated context, RAG, memory systems (neural memory architectures and production memory), agent communication (open protocols like MCP and A2A), tool use, evaluation (long-context benchmarks such as LongBench, RULER and InfiniteBench) and applications (automated scientific research systems and enterprise production systems).
032026 Agent Era update: from Context Engineering to Agent Engineering
Since March 2026 the repo has been reorganized around a shift in gravity: from 'how to pack the best prompt' to 'how agent systems manage runtime state, memory, tools, protocols and long-horizon execution'. It added sections on agent harnesses/runtimes (planning, subagents, checkpoints, sandboxes, human approvals), open protocols, coding agents and project memory, and observability — stating plainly that context engineering now sits inside a larger stack and, in practice, is becoming agent engineering.
04The formal definition and 'why it matters'
The repo gives a rigorous definition: context is not the single prompt a user sends, but the complete information payload provided to the LLM at inference time (instructions, knowledge, tools, memory, state, query). It models context engineering mathematically as an optimization problem over context-assembly functions, and a comparison table shows prompt engineering as a static string versus context engineering as a dynamic multi-component system. It also argues the motivation: ambiguous human intent, stale knowledge, hallucination and provenance problems all demand systematic context management.
05How to use it to go deep on context engineering
Use it as a map: to follow the 2026 agent shift, start with the agent harnesses and open protocols sections; to build foundations, read the definition and 'why context engineering' sections; to work on long-context evaluation, visit the Evaluation section for LongBench, RULER, InfiniteBench and observability practices. Every entry links out, so you can go deep on any paper or tool, and reading it alongside the companion survey gives the full picture.

Practice tips

Start with the 2026 Agent Era Update in the README to see which section your interest falls into, then decide where to begin.
Pick one category (e.g. Memory Systems) and read 2-3 of the listed papers closely, noting how they converge on the same problem.
Copy the prompt vs context engineering comparison table into your notes as an anchor for future judgment.
Read your chosen section alongside the companion survey paper and summarize the state of that direction in your own words.