Context Engineering
An engineering practice that optimizes AI model output by proactively supplying and organizing contextual information, such as project specifications, domain knowledge and tool configuration. It is a key method for improving the effectiveness of AI coding assistants.
Core Facts
AI development has evolved through four stages: single large model calls, prompt engineering, agents and tool calls, and loop engineering.
大多数 AI 编程智能体无法有效完成复杂任务的主要原因是上下文缺失，而非模型能力不足
Context Engineering is a new paradigm that provides AI with a complete, systematic context, including project documentation, code examples, coding rules, development patterns, and validation mechanisms.
Timeline (last 90 days)
A large portion of long-task agent failures stem from improper context management, for example, unbounded accumulation of historical messages that dilutes key information, or loss of constraints that must be followed when compressing summaries.
Each of the three loop layers depends on a reliable context supply: the execution layer needs the current task state, the verification layer needs the original requirements to compare against the output, and the improvement layer needs complete traces.
Anthropic technology team member Lamis systematically reviewed the evolution path of Context Engineering over the past year at AI DevCon.
Context engineering treats all information visible to the model (system prompts, retrieved snippets, tool outputs, conversation history, user profiles) as a resource system that requires proactive design and dynamic governance.
Whether knowledge should be placed in context or solidified into the model through fine-tuning is still mainly decided by empirical judgment and small-scale comparative experiments, and no general decision framework exists yet.
The ownership of context engineering (implementation at the application layer vs. pushing down to the model or framework layer) currently has blurred boundaries, and both routes have been attempted.
Typical patterns in context engineering practice include layered context, dynamic summarization, tool output filtering and context budgeting.
More context is not better; more relevant context is better, and the performance bottleneck of multi-agent systems often lies not in model capability but in handoff interface design.
Context engineering refers to the process of actively designing, filtering and organizing input context for language models or agent systems, as distinct from guiding models solely through prompt wording.
In multi-agent workflows, giving the next agent too much context at handoff causes it to drown in irrelevant history and even treat outdated facts as evidence, while giving too little forces it to guess blindly or ask for help at the cost of reopening the entire history.
40 more timeline events
All Facts (20)
AI development has evolved through four stages: single large model calls, prompt engineering, agents and tool calls, and loop engineering.
90%Verified大多数 AI 编程智能体无法有效完成复杂任务的主要原因是上下文缺失，而非模型能力不足
80%VerifiedContext Engineering is a new paradigm that provides AI with a complete, systematic context, including project documentation, code examples, coding rules, development patterns, and validation mechanisms.
80%VerifiedContext Engineering focuses on how to precisely manage the context window fed into the model, Harness Engineering leans toward the infrastructure layer, and both appeared around the same time as Loop Engineering.
65%VerifiedThe system prompt is the highest-priority instruction layer in large language model conversation architecture, injected before user messages and influencing how the model processes all subsequent inputs.
65%VerifiedContext engineering is the systematic design and optimization of the contextual information fed to an LLM, enabling the model to obtain the most relevant information within a limited context window to complete a task.
65%VerifiedContext engineering refers to the engineering practice of systematically designing, managing, and optimizing the contextual information fed into large language models, which differs from prompt engineering in that the latter focuses on wording optimization for a single interaction.
65%Unverified吴恩达认为过去几年行业注意力集中在提示工程和上下文工程上，但对于需要持续运行数天甚至数周的Agent来说这远远不够
90%Unverified在真实上下文工程场景中，上下文散布在多个来源：本地文件系统、数据库、Web、长期记忆等，每种上下文源需要对应的搜索工具
90%UnverifiedAGENTS.md的实践与'上下文工程'（Context Engineering）的理念密切相关，代表了从提示工程到上下文工程的演进
85%Unverified上下文工程的重心已从'如何塞进更多信息'转变为'如何确保正确的信息出现在正确的位置'
80%Unverified在AI编程中，提前设计好数据结构并提供JSON文档给AI（即'上下文工程'）能大幅提升代码生成的准确度，这是区分AI编程新手与熟练用户的核心能力差异之一
75%Unverified行业正在经历从'模型中心'到'系统中心'的范式转移，单纯提升模型参数和训练数据的边际收益在递减
70%UnverifiedContext Engineering is defined as a systematic methodology that goes beyond prompt engineering, with prompt engineering being a subset of context engineering.
60%UnverifiedContext engineering is a further evolution beyond RAG, focusing not only on recall rate of information but also on the structural quality, credibility levels, and conflict marking of the injected context.
60%Unverified上下文工程使得零编程基础的用户也能开发复杂项目
60%UnverifiedA large portion of long-task agent failures stem from improper context management, for example, unbounded accumulation of historical messages that dilutes key information, or loss of constraints that must be followed when compressing summaries.
50%UnverifiedEach of the three loop layers depends on a reliable context supply: the execution layer needs the current task state, the verification layer needs the original requirements to compare against the output, and the improvement layer needs complete traces.
50%UnverifiedThe ownership of context engineering (implementation at the application layer vs. pushing down to the model or framework layer) currently has blurred boundaries, and both routes have been attempted.
50%UnverifiedContext engineering treats all information visible to the model (system prompts, retrieved snippets, tool outputs, conversation history, user profiles) as a resource system that requires proactive design and dynamic governance.
50%