Compare context engineering and prompt engineering: scope, artifacts, skills, and when each applies. Why prompting became a subset in the agent era.
Ka Ling Wu
Co-Founder & CEO, Upsolve AI
10 min
Prompt engineering is the craft of writing one good instruction.
Context engineering is the process of controlling everything the model sees at inference time: instructions, retrieved documents, tool definitions, memory, and conversation history.
As AI systems shifted from single-turn chatbots to long-running agents, prompt engineering became a component of context engineering.
Context Engineering vs Prompt Engineering: The Comparison
Prompt Engineering | Context Engineering | |
Scope | The instruction text in a single request | Everything in the context window, on every inference call |
Artifacts | System prompts, templates, few-shot examples | Retrieval pipelines, memory stores, tool schemas, compaction logic, semantic layers |
Time horizon | One turn | Hundreds of turns across an agent loop |
Core skill | Writing clear, unambiguous instructions | Information architecture: deciding what enters, stays in, and leaves the window |
Failure mode | Vague or conflicting instructions | Context rot, poisoning, token bloat, lost-in-the-middle degradation |
Optimized by | Rewording, restructuring, adding examples | Retrieval quality, compression, isolation, just-in-time loading |
Relationship | A subset, one component of the assembled context | The superset discipline that contains prompting |
When it's enough | Single-turn tasks, chatbots, one-shot generation | Agents, RAG systems, anything that runs longer than one exchange |
If you only remember one row, make it the last two. Prompt engineering is one operation inside a larger system. Context engineering is the system.
Key Takeaways |
|---|
|
What Is Prompt Engineering?
Prompt engineering is the practice of writing and structuring the instruction a language model receives to maximize the quality of a single response. It covers phrasing, output formatting, role assignment, chain-of-thought scaffolding, and few-shot examples. The unit of work is one request; the feedback loop is one response.
For roughly two years, this was the whole game. If GPT-4 gave you a mediocre answer, you rewrote the prompt. Entire job listings, courses, and a small library of "500 best prompts" ebooks were built on that loop.
What Is Context Engineering?
Context engineering is the practice of curating everything that lands in a model's context window at inference time: system prompts, user input, retrieved documents, tool definitions, conversation history, and long-term memory; so the model has exactly what it needs and nothing it doesn't. Andrej Karpathy, who popularized the term in a June 2025 post, called it the "delicate art and science of filling the context window with just the right information for the next step."
Shopify CEO Tobi Lütke argued the term simply describes the real skill better, "the art of providing all the context for the task to be plausibly solvable by the LLM", and Simon Willison predicted it would stick, because unlike "prompt engineering," people's inferred definition of context engineering matches what practitioners actually do. By early 2026, Gartner had issued its own enterprise definition, which is usually the moment a practitioner term stops being a trend.
Why Agents Created the Shift
Nothing about prompt engineering broke. What broke was the assumption that one well-written block of text could carry a system that runs for hours.
Chatbots Answer; Agents Loop
A chatbot handles one exchange: question in, answer out, context discarded. An agent runs a loop (plan, call a tool, read the result, revise the plan, call another tool) and every iteration adds tokens. Anthropic's engineering team describes production agents engaging in conversations spanning hundreds of turns, requiring careful strategies for what to keep, summarize, or store in external memory as the window fills.
Bigger Windows Didn't Solve It
Just use a million-token window and stuff everything in.This turned out to be a trap. Chroma's research on context rot tested 18 models, including GPT-4.1 and Claude 4, and found performance degrades non-uniformly as input length grows, even on tasks as simple as repeating a word. Models pay attention unevenly across long inputs; low-signal tokens actively dilute the high-signal ones. Our context rot guide covers the benchmark evidence in depth.
This means context is like a budget. Anthropic's guiding principle for agent design is finding the smallest set of high-signal tokens that maximizes the likelihood of the outcome you want. That's an editorial judgment applied at system scale, which is why the discipline needed a name of its own.
Production Teams Learned It the Hard Way
The Manus team rebuilt their agent framework four times before landing on patterns that held up: design around KV-cache hit rates, mask tools instead of removing them, externalize memory to the file system, and, counterintuitively, preserve the agent's error messages in context so it learns from failed attempts instead of repeating them.
Notice that none of those patterns involve wordsmithing. That's the tell. When the leverage moved from phrasing to plumbing, the field needed a new word.
Key insight: Prompt engineering optimizes what you say to the model. Context engineering optimizes what the model knows when you say it. In a one-turn system those are the same thing. In an agent, they diverge fast.
The Four Operations of Context Engineering
LangChain's widely-adopted taxonomy breaks the discipline into four operations and the first one is where prompt engineering lives. We apply the same four operations to data agents in our agentic context engineering guide.
Write
Save context outside the window so it survives. Agents jot their plan to a scratchpad and persist notes to memory files, so that critical information outlives window truncation instead of vanishing when the conversation is compressed. Claude Code's to-do lists and CLAUDE.md-style project files are this operation in the wild. Classical prompt authoring (the system prompt, tool descriptions, few-shot examples) is the static context these operations then manage.
Select
Pull in the right dynamic information at the right moment. RAG is the best-known select technique, but the frontier in 2026 is just-in-time retrieval: instead of pre-loading everything, the agent holds lightweight references (file paths, queries, links) and loads content only when a step requires it. Anthropic calls this progressive disclosure, and it's the reason modern agents navigate large codebases without swallowing them whole.
Compress
Reduce tokens without losing signal. Compaction summarizes a conversation nearing the window limit and reinitializes with the summary; pruning drops stale tool outputs. Done badly, compression deletes the one detail the agent needed forty turns later, which is why teams treat their summarization prompts as production code.
Isolate
Split work across sub-agents with separate, clean windows. Anthropic's multi-agent research system uses an orchestrator that dispatches focused sub-agents, each doing deep work in its own context and returning a distilled summary. Isolation prevents one task's noise from poisoning another's reasoning.
When Prompt Engineering Is Still the Right Tool
If you're doing any of the following, prompt engineering alone is the correct scope:
Single-turn generation: Drafting emails, summarizing a pasted document, reformatting data. There's no loop, so there's no context to manage beyond the prompt itself.
Stable, repeated tasks: A classification prompt that runs a million times a day benefits enormously from careful wording and examples, and not at all from retrieval infrastructure it doesn't need.
Prototyping: Before you build a memory system, prove the model can do the task with ideal context pasted in by hand. If it can't, no pipeline will save it.
The moment your system makes decisions about what the model sees (retrieval, memory, tool results, history) you've crossed into context engineering, whether you call it that or not. Until then, you're prompting, and that's fine.
What This Means If Your Agent Answers Data Questions
An analytics agent that answers "why did churn spike in March?" against a live database is a context engineering problem in its purest form. The model doesn't need a cleverer prompt. It needs to know your schema, your metric definitions, which "revenue" column is the real one, and what this specific user is allowed to see.
That's exactly what a semantic layer is: pre-engineered context for data questions. It's the "select" and "write" operations done once, correctly, so every query the agent handles starts from trusted definitions instead of guessing at column names. Teams that skip this step get agents that hallucinate joins; teams that do it get answers users can actually trust.
Upsolve's embedded analytics agents sit inside your SaaS product and answer your customers' natural-language data questions. The reason the answers are trustworthy is the context architecture underneath: a governed semantic layer, per-tenant security context, and dashboards the agent grounds its answers in. You could spend two quarters building that context pipeline yourself, or embed it in weeks.
Frequently Asked Questions
What is the difference between context engineering and prompt engineering?
Prompt engineering optimizes the instruction text sent to a model in a single request. Context engineering manages everything in the model's context window across an entire session: instructions, retrieved data, tool definitions, memory, and history. Prompt engineering is a subset of context engineering: one component of the assembled context.
Did context engineering replace prompt engineering?
No, it absorbed it. Writing clear instructions is still a required skill; it's simply no longer sufficient for agents and RAG systems where most context is assembled programmatically. For single-turn tasks, prompt engineering alone is still the right scope.
Who coined the term context engineering?
Andrej Karpathy popularized it in June 2025, defining it as filling the context window with just the right information for the next step. Shopify CEO Tobi Lütke and Simon Willison amplified the framing, and Anthropic's "Effective context engineering for AI agents" became the discipline's reference document.
Is context engineering just RAG?
No. RAG is one "select" technique within a broader toolkit that also includes writing static context, compressing history through summarization, isolating work across sub-agents, and managing long-term memory. RAG decides what to retrieve; context engineering also decides what to keep, compress, and discard.
How long does it take to learn context engineering?
If you already prompt well, the core four operations take days to understand. Production competence takes longer, because the hard part is debugging silent failures like context poisoning, which only shows up when you build and instrument a real system. Expect weeks of hands-on iteration, not a weekend course.
Do larger context windows make context engineering unnecessary?
The opposite. Research on context rot shows model accuracy degrades as inputs grow, even within the supported window. Bigger windows raise the ceiling on what you can include, which makes deciding what you should include more important, not less.
Sources
Anthropic Engineering: Effective context engineering for AI agents; the discipline's reference document, including the smallest-high-signal-tokens principle. https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents
Anthropic (Claude Blog): The new rules of context engineering for Claude 5 generation models; the 80% system prompt reduction and the shift from rules to judgment. https://claude.com/blog/the-new-rules-of-context-engineering-for-claude-5-generation-models
Andrej Karpathy (X): June 2025 post popularizing the term "context engineering." https://x.com/karpathy/status/1937902205765607626
Simon Willison: Context engineering write-up quoting Tobi Lütke's framing; why the term's inferred definition matches the real skill. https://simonwillison.net/2025/Jun/27/context-engineering/
LangChain: The rise of context engineering; Harrison Chase's definition and the "prompt engineering is a subset of context engineering" framing. https://www.langchain.com/blog/the-rise-of-context-engineering
LangChain: Context engineering for agents; the write/select/compress/isolate strategies, scratchpads, and memory. https://www.langchain.com/blog/context-engineering-for-agents
Chroma Research: Context rot: how performance of 18 models degrades with input length, including on a repeated-words task. https://research.trychroma.com/context-rot
Drew Breunig: How long contexts fail: context poisoning, distraction, confusion, and clash. https://www.dbreunig.com/2025/06/22/how-contexts-fail-and-how-to-fix-them.html
Manus: Context engineering for AI agents: lessons from four framework rebuilds. https://manus.im/blog/Context-Engineering-for-AI-Agents-Lessons-from-Building-Manus
Anthropic Engineering: How we built our multi-agent research system; orchestrator/sub-agent isolation and long-horizon conversation management. https://www.anthropic.com/engineering/multi-agent-research-system
Upsolve.ai: Embedded analytics agents grounded in a governed semantic layer. https://upsolve.ai/embedded-bi
Try Upsolve for Embedded Dashboards & AI Insights
Embed dashboards and AI insights directly into your product, with no heavy engineering required.
Fast setup
Built for SaaS products
30‑day free trial
