Answer
What is context engineering?
Deciding what a system should be holding at each step of a job, rather than how to word the request.
It is the practice of deciding which information should be present while a system works, and keeping that set small and relevant. Prompt engineering is about wording one request; this is about managing what accumulates across a job.
The term marks a shift in where the difficulty sits. When the interaction was one request and one reply, the wording of the request was most of the design. Once a system works across many turns, calling tools and reading results, the wording is a small part of what determines the outcome and the composition of everything present becomes the larger part: instructions, tool definitions, retrieved documents, prior messages, results of earlier calls.
The central finding, and the one that surprises people, is that this is a budget rather than a container. Adding more relevant material improves results up to a point and then stops, and past that point additional volume reduces retrieval accuracy — the ability to use a specific fact that is genuinely present. So the objective is not to supply everything that might bear on the task. It is to find the smallest set of high-signal material that makes the right outcome likely, which is an editorial problem rather than a capacity one.
Four techniques do most of the work in practice. Loading material at the moment it is needed rather than up front, so that identifiers and paths are held and content is fetched when a step requires it. Summarising and restarting when a session approaches its limit, carrying forward the summary rather than the history. Writing notes to durable storage outside the session, so a fact survives the window it was learned in. And handing focused sub-tasks to separate runs with clean context, which return a condensed result rather than their working.
Each of these is a trade rather than a free improvement, and the trades are worth naming. Loading on demand costs a round trip and risks the agent not fetching something it needed. Summarising loses detail, and which detail is lost is not chosen by anyone. Notes go stale. Sub-tasks return a summary written by a party that saw more than you did. The discipline is knowing which loss you are accepting, not avoiding loss.
For an ordinary business the useful translation is short. Supply what bears on this job rather than everything about the subject. Keep durable facts in a file rather than in a conversation. Break long work into stages that each write down their result. And when a session has been running a while and the answers are getting vaguer, start a new one from the written state rather than pressing on, because the accumulated material is now working against the task rather than for it.
Prompt engineering asks what should I say; context engineering asks what should be in front of it when it decides, which is the harder and more consequential question.
Siddharth Sharma, Context Theory
Related questions
Is this just a new name for prompt engineering?
It covers a different object. Prompt engineering concerns how one instruction is written; this concerns what is present while a multi-step job runs, most of which arrives without anyone writing it — tool results, retrieved files, earlier turns. The reason the distinction earns a name is that the second is where multi-step systems actually fail, and no amount of attention to the first addresses it.
Does this matter for someone who is not building agents?
The budget property does. Anyone who has watched a long conversation get progressively less precise has met it, and the remedy is the same at every scale: state the job again with only the material that bears on it, in a new session, rather than continuing to add to one that has accumulated everything.
METHOD
Every figure below carries its source and the date it was verified. Nothing on this page is asserted.
The numbers on this page.
| What | Value | Specific to |
|---|---|---|
| Visibility lift in AI-generated answers from GEO methods | up to 40% | Category-wide |
| AI-cited sources that also rank in the Google organic top 10 | 10% | Category-wide |
Aggarwal et al., "GEO: Generative Engine Optimization", Princeton / Georgia Tech / IIT Delhi / Allen Institute for AI — KDD 2024 · GEO-bench · 10,000 queries across 8 domains · verified
2026 generative engine citation study · fewer than · verified
What is specific to this page.
| Kind | Claim | Check it against |
|---|---|---|
| Software | Context engineering is defined as curating and maintaining the set of information present during inference, which is distinct from prompt engineering because most of what is present in a multi-step run — tool results, retrieved documents, prior turns — arrives without being authored. | Anthropic's engineering write-up on context engineering for agents, which draws this distinction explicitly. |
| Workflow | Retrieval accuracy declines as the volume of held information grows, so the objective is the smallest set of high-signal material that makes the desired outcome likely rather than the most complete set available. | Published measurements of retrieval accuracy against context length, and the same test run on a specific fact placed in a short and a long context. |
| Response | Four named techniques address the budget: loading material at the point of need, summarising and reinitialising near the limit, writing notes to storage outside the session, and delegating focused work to runs with clean context that return condensed results. | The same engineering write-up, which names just-in-time retrieval, compaction, structured note-taking and sub-agent architectures. |
| Buying behaviour | Each technique trades a specific loss rather than removing one: on-demand loading risks an unfetched dependency, summarisation discards detail nobody selected, notes go stale, and delegated summaries are written by a party that saw more than the reader. | Comparing a compacted session's summary against the discarded history for facts that did not survive. |
Each row would be wrong on another industry's page. Where a sourced figure exists it is in the table above instead; these are the constraints that shape the work and do not happen to be numbers.
Start with the measurement.
Reading about a benchmark is not the same as knowing your own number. The audit produces yours, measured rather than estimated.
$497 · delivered in 5 business days · credited against month one