From Prompt Engineering to Context Engineering: The Quiet Discipline Shift of 2026

For two years we argued about adjectives. Now we're arguing about architecture.

If you build with LLMs, you've felt it. The vocabulary shifted under your feet sometime in the last nine months and most people didn't notice. "Prompt engineer" — a job title that briefly ruled LinkedIn — has gone quiet. "Context engineering" has taken its place. The shift is small in words and enormous in what it implies about how serious teams actually build agents.

The Prompt Was Never the Product

In 2023 and 2024, a lot of "AI engineering" really was prompt engineering. You had one model, no tools, no memory, no retrieval. The whole interaction fit in a single chat window. If the output wasn't right, the answer was usually a better instruction: more specific, more examples, more structure. You could get a long way on cleverness alone.

Then the models got tool use. They got memory. They got retrieval over private data. They got multi-step plans. Suddenly the prompt wasn't the whole system — it was a tiny invocation at the center of a much larger machine. And a tiny, clever instruction at the center of a noisy machine is, at best, a small part of the answer.

Sourcegraph put it well: the clearest tell that you've crossed from prompt engineering into context engineering is whether your improvements come from rewording or from rewiring.

What Context Engineering Actually Is

Context engineering is the discipline of deciding what the model sees, when, and in what shape. It's plumbing. It's the unsexy work that determines whether an agent actually does the thing you wanted or hallucinates confidently about the wrong file.

In practice it breaks into a handful of questions every agent team has to answer:

• What do we retrieve? The right snippet, the right document, the right memory entry — not the whole corpus.
• When do we fetch it? Before the call, during, lazily after an error?
• How do we compress it? Long histories blow up the context window and drown the signal. Summarization, eviction, and selective replay become first-class features.
• What do we forget? Memory is not a log. Deciding what not to remember is often more important than what to store.
• What do we put back? Tool outputs, intermediate results, scratch state — all of it competes for the same finite window.

None of this shows up in the prompt. All of it shows up in whether the agent works.

The Tell in Production

You can see the difference in the failure modes. Prompt-engineered systems fail because the model misunderstood the instruction — and the fix is a paragraph of carefully chosen words. Context-engineered systems fail because the model saw the wrong information, the wrong version, or stale data — and the fix is a different data path, a different retrieval ranker, a different compaction policy.

The first kind of bug lives in a markdown file. The second kind lives in the system architecture. Which is why "context engineering" is being staffed by people who used to be backend engineers, not people who used to be copywriters.

How We Think About It

At heimdall.engineering, our agent layer is mostly context engineering at this point. The system prompt is small and stable; the interesting work is in what the agent pulls in. Which files does it see when it opens a task? Which memories from past sessions are relevant? Which tool outputs are summarized and which are kept verbatim? Which errors are worth replaying and which are noise?

We spend far more time on retrieval ranking, memory eviction, and tool-output compression than we do on the prompt itself. When something goes wrong, it's almost never "the model didn't understand." It's "the model saw the wrong thing." Same observation every serious agent team is reporting in 2026.

Why This Is the Real Engineering Shift

Prompt engineering was the discipline of talking to a model. Context engineering is the discipline of building the world the model operates in. The first is a writing skill. The second is a systems skill.

And once you accept that, a lot of downstream questions resolve themselves. Why are agent frameworks getting more opinionated about memory? Why is everyone obsessed with retrieval evals? Why did "RAG" stop being a buzzword and start being a product surface? It's all context engineering. It's all plumbing. And plumbing, as it turns out, is where the leverage is.

If you're hiring for agent work in 2026, stop interviewing for prompt cleverness. Interview for someone who can design a retrieval pipeline, reason about eviction policy, and debug why the model saw a three-hour-old version of a file. The discipline changed. The job description should too.