The main claim: this is not new
Most writing on context engineering dates it to 2025, when the term became popular for AI agents. The paper argues the problem is older. Ever since computers had to respond to people, someone had to decide what information about the situation the system could use. The paper builds its definition on a classic one from human-computer interaction research: context is “any information that can be used to characterize the situation of entities that are considered relevant” to an interaction. Context engineering then is “the systematic process of designing and optimizing context collection, storage, management, and usage.”
The four eras
| Era | Period | What the machine can do | Who does the context work |
|---|---|---|---|
| 1.0 Primitive computation | 1990s to 2020 | Structured input, sensor signals, rules | Designers act as “intention translators” |
| 2.0 Agent-centric intelligence | 2020 to now, starting with GPT-3 | Reads natural language, handles some ambiguity | Developers and users curate prompts, memory, tools |
| 3.0 Human-level intelligence | Future | Reads high-entropy signals such as emotion and social cues | Mostly the system |
| 4.0 Superhuman intelligence | Speculative | Builds context itself and surfaces needs people have not stated | The machine |
The thread through all four is what the paper calls entropy reduction. People fill in missing context without noticing; machines cannot. So someone has to turn high-entropy intent into low-entropy, machine-usable form. The smarter the machine, the less of that work falls on the human. Read this way, “2.0” is not a version of a tool but a point in that curve.
The design considerations
The paper groups the practical part into four stages:
- Collection and storage. Two principles: minimal sufficiency, “The context value lies in sufficiency, not volume”, and semantic continuity, keeping meaning rather than raw data.
- Management. Timestamps, tagging, hierarchical notes, layered memory, isolating sub-agents and lightweight references instead of full content. Short-term memory is context that is relevant now; long-term memory is context important enough to be abstracted and kept, with a transfer step between them.
- Usage. Selecting what to load by semantic relevance, logical dependency, recency, frequency and user preference, which the paper calls “attention before attention”. Sharing context between agents, within one system through prompts, fixed-schema messages or shared memory, and across systems through adapters, shared formats, readable summaries or vectors.
- Self-baking. The agent digests its own history into lasting structures: natural-language summaries, facts in a fixed schema, hierarchical memory or compressed embeddings. Each trades off structure, readability and upkeep.
It ends with open problems: storage, degradation on long inputs, errors that spread through memory, evaluation, and a proposed “Semantic Operating System for Context” in which context “is actively managed and evolved.”
The common misunderstanding
The title invites the reading that 2.0 is a new method that replaces something called 1.0, the way people speak of “prompt engineering 2.0”. It is not. The paper offers no new technique, benchmark or code. Its value is a vocabulary and a frame: what counts as context, why it costs effort, and which choices every system makes about collecting, storing, selecting and sharing it. If you want the techniques, the July 2025 survey by Lingrui Mei and colleagues is the broader map, and posts such as Anthropic’s are more hands-on; see What does Anthropic say about context engineering?
What this means for a team
Two ideas from the paper translate directly to a team using several AI tools. First, cross-system sharing: the paper lists the trade-offs between schemas (precise but needing coordination), summaries (flexible but harder to parse) and vectors (compact but hard to inspect). That is the choice a team faces when knowledge has to move between ChatGPT, Claude and a coding agent. Second, self-baking: a team whose AI only recalls old chats keeps repeating itself; one whose corrections become lasting rules accumulates knowledge. For the agent-level version of that loop, see What is agentic context engineering (ACE)? and What is a memory OS for AI agents?
Where Bombos fits
Bombos does the cross-system part for a team: one plugin that gives every AI tool the team uses the same memory, with team knowledge, learned rules and what is already in mail, documents and repos. Corrections can be kept as learned rules for later work, which is close to what the paper calls self-baking, and you can see what each answer rests on. It runs in our cloud or on your own server.
Related questions
Who wrote Context Engineering 2.0?
Qishuo Hua, Lyumanshan Ye, Dayuan Fu, Yang Xiao, Xiaojie Cai, Yunze Wu, Jifan Lin, Junfei Wang and Pengfei Liu. It was submitted to arXiv on October 30, 2025 as 2510.26493.
Is Context Engineering 2.0 a new technique or framework?
No. It is a conceptual paper: a definition, a history and design considerations. It does not ship code or a benchmark.
What is self-baking in the paper?
The paper's term for an agent digesting its own context into lasting knowledge structures, such as summaries, extracted facts or hierarchical memory. Its line: "Without self-baking, agents simply recall; with it, they accumulate knowledge."
How is it different from "A Survey of Context Engineering for Large Language Models"?
The survey by Lingrui Mei and colleagues (July 2025) maps techniques from over 1,400 papers. Context Engineering 2.0 is shorter and conceptual: it asks what context is and how the discipline evolved.
