Glossary

Context engineering: what it is and why it replaced prompt tuning

Context engineering is the practice of deciding exactly what information a language model sees at each step of a task: its instructions, examples, retrieved data, tool results and memory. It grew out of prompt engineering as AI moved from single questions to agents that run many steps. Most agent failures in production trace back to context the model was missing or context it should never have had.

Shaun DevanFeaturing insights fromShaun Devan · Founder & CEO

Updated · 3 min read

Get in touchSee our work

Context engineering vs prompt engineering

Prompt engineeringContext engineering
FocusWording of one instructionEverything the model sees, at every step
ScopeA single callA multi-step agent run
IncludesPhrasing, examplesRules, retrieved data, tool outputs, history, memory
Main riskA vague instructionToo little, too much, or stale information

What good context looks like

  • The business rules the task depends on, written plainly, not left for the model to infer.
  • Only the documents and records relevant to this case, not the whole archive.
  • Tool results trimmed to what the next step needs.
  • A few real examples of hard cases and the right outcome.

Why it is an engineering job

Context is assembled by code: retrieval, database queries, summaries of earlier steps. It has to be tested like code too, with evals that show whether a change to what the model sees improves the result. Native's engineers treat it as part of the build, alongside integrations and evals.

Sources

Frequently asked questions

Is context engineering the same as prompt engineering?

It is the broader discipline. Prompt engineering tunes one instruction; context engineering designs everything the model sees across a whole agent run.

Why do AI agents need context engineering?

Agents take many steps and gather information as they go. Without deliberate design, they lose track of the rules, drown in irrelevant data or act on stale results.

Does a bigger context window remove the need for context engineering?

No. Models still perform better with focused, relevant context, and more tokens cost more and take longer.

Keep reading

GlossaryRetrieval-augmented generation (RAG): what it isRetrieval-augmented generation (RAG) finds relevant passages in your own data and gives them to a language model, so its answers draw on your facts.GlossaryModel Context Protocol (MCP): what it is and why it mattersThe Model Context Protocol is an open standard for connecting AI models and agents to tools, data and business systems through one common interface.GlossaryAI evals: how AI systems get testedAI evals are test sets of real cases with known right answers, run against an AI system before release and after every change to measure its quality.InsightThe AI engineering skills your team needs, from the production floorFour skills decide whether a team ships AI that works: evals, software fundamentals, coding agents and shaping the build, as they look in production.NativeNative Agents

Put AI to work across your whole business.

Bring the hardest problem on your list. In 30 minutes we'll show you where AI will have the most impact first, and what it takes to get there.

Book a callSee our work