Context Engineering 2.0: 30 Years From Context-Aware to Context-Cooperative Systems

Forum topic · ✨步子哥 · 2025-11-28

Summary

This visual essay presents Context Engineering 2.0, framing the discipline as a 30-year evolution from context-aware to context-cooperative systems and as an entropy-reduction process that bridges the cognitive gap between humans and machines. It traces the field from Bill Schilit's 1994 context-aware computing concept and Anind Dey's 2001 definition through the Context Toolkit framework, then proposes a systematic framework defined as Context Engineering = Collection x Management x Usage. The post contrasts Era 1.0 (passive response, structured data, single sensors, likened to an infant eating only pureed food) with Era 2.0 (active collaboration, raw information handling, multimodal fusion, likened to an adult eating steak). Key techniques covered include multi-modal context collection, hierarchical memory architectures, sub-agent context isolation in tools like Claude Code, and lightweight external referencing instead of bloating the context window. The piece introduces Self-Baking as the watershed from tool to partner, separating storage from learning so AI accumulates knowledge rather than only recalling prior inputs.

Overview

Context Engineering 2.0: The Context of Context Engineering frames the discipline as a 30-year evolution from *context-aware* to *context-cooperative* systems. The authors define context engineering as an entropy-reduction process that bridges the cognitive gap between human and machine information handling:

> Cognitive Gap = Human context-processing capacity − Machine context-processing capacity

30-Year Timeline

1994 — Bill Schilit first proposes the concept of *context-aware computing*.
2000 — Anind Dey's team develops the Context Toolkit framework.
2001 — Anind Dey publishes the still widely cited formal definition of context.
Era 1.0: Context-Aware — Machines behave like infants that can only ingest *pureed food* (structured data).
Era 2.0: Context-Cooperative — Machines behave like adults that can directly digest a *steak* (raw, unstructured information).

Systematic Framework

\text{Context Engineering} = \text{Collection} \times \text{Management} \times \text{Usage}

Collection

Upgrade from single sensors to multi-modal fusion.
Capture environment, location, identity, activity, and other situational signals.

Management

Hierarchical memory: short-term and long-term stores.
Sub-agent isolation: Claude Code spawns sub-agents so each executes tasks with its own clean context.
Lightweight referencing: keep large files outside the context window and store only a *pointer* inside it.

Usage

Shift from *passive response* to *active collaboration*.
Improve high-entropy context consumption: from processing only "preprocessed food" to digesting raw source material.

Era 1.0 vs Era 2.0

| Era 1.0 | Era 2.0 | |---|---| | Passive response | Active collaboration | | Structured data only | Raw information capable | | Single sensor | Multi-modal fusion |

Example of Era 2.0 collaboration: While you are writing a paper, the system analyzes your writing progress, detects that you are stuck on Chapter 3, and proactively asks, "Want me to help you reorganize the logic?" You agree, it drafts an outline, you revise it, and it adjusts based on your feedback.

Self-Baking: From Quantitative to Qualitative Change

Self-Baking separates storage from learning.

Without Self-Baking — the AI only *recalls* ("What did you say last time?").
With Self-Baking — the AI *accumulates knowledge* ("I know what you like.").

The authors call this the watershed between *tool* and *partner*. They argue context engineering must move from quantitative accumulation to qualitative change, with the distilled core summary fused into model parameters themselves.

Resources

GitHub: <https://github.com/GAIR-NLP/Context-Engineering-2.0>
Paper PDF: <https://arxiv.org/pdf/2510.26493>
Slides (Quark Pan): <https://pan.quark.cn/s/52ad1ae833a7>