The agent that got dumber as I fed it
Early this year I spent an afternoon making an agent worse by improving it. The system — a document-heavy assistant on this project — kept missing details, so I did the obvious thing and gave it more: more retrieved passages, more conversation history, more of the internal wiki. Every addition felt safer. The answers got vaguer instead, and somewhere along the way the agent began confidently ignoring an instruction that sat in plain view, buried in the middle of its own context. I had been treating the context window as a bucket, where pouring in more loses nothing. That afternoon is why I stopped rolling my eyes at the most contested term of the past year.
A term coined in a week
“Context engineering” went from tweet to canon in about ninety days. On June 19, 2025, Tobi Lütke wrote that he preferred it to prompt engineering because it “describes the core skill better: the art of providing all the context for the task to be plausibly solvable by the LLM.” Within a week, Andrej Karpathy seconded it: “the delicate art and science of filling the context window with just the right information for the next step.” Simon Willison, who had spent years defending the older term, conceded within days that the new one would fare better because its plain reading matches what practitioners actually do.
My own reaction at the time was less generous. I have watched enough engineering buzzwords arrive with a manifesto and leave without an obituary, and every instinct said this was prompt engineering with a fresh coat of paint — the same renaming reflex that turns everyone who writes SQL into a data scientist.
The evidence predates the label
Here is what changed my mind: the finding came before the name. In the Lost in the Middle study, published in TACL in 2024, Liu and colleagues showed that language models reliably use information at the beginning and end of a long context and degrade significantly when the relevant material sits in the middle. That paper predates the term entirely. A rebrand renames a practice; this named a measurement.
The measurements kept coming. In July 2025, Chroma’s Context Rot report evaluated 18 models and found that performance degrades as input length grows even on trivial tasks — and that on a long-memory benchmark, a focused prompt of roughly 300 tokens outperformed a roughly 113,000-token prompt containing the same relevant information. Chroma sells retrieval infrastructure, so weight that accordingly. The peer-reviewed spine holds without them: an ICLR 2025 study found that for many long-context models, adding retrieved passages improves output only up to a point, after which quality declines, with irrelevant-but-plausible passages doing much of the damage. More context is not monotonically better, and that single sentence retires the bucket.
The physics under the buzzword
Anthropic’s engineering post from September 2025 gives the cleanest account of why. Context, it argues, must be treated as “a finite resource with diminishing marginal returns”, because attention is pairwise: every token in the window relates to every other token, so the load grows with the square of the input while the model’s capacity to track it does not. The post calls this the model’s attention budget, and the phrase is the whole discipline in two words.
I will note the incentives, because they are everywhere. The loudest authors of this canon all sell a remedy — LangChain a framework, Chroma a database, Anthropic the models themselves. That is precisely why the peer-reviewed results matter: the claim survives the sellers. The techniques that follow from it are unglamorous and concrete — compact a long conversation into a summary and start fresh, retrieve documents at the moment they are needed instead of preloading them, and split work across sub-agents that each get a clean window and report back in condensed form.
Run it like a budget
The frame that made it click for me is managerial. LangChain’s taxonomy sorts every technique into four moves — write, select, compress, isolate — and all four are allocation decisions. Every token in the window competes with every other token for a fixed pool of attention, which means the job is not gathering information but rationing it, and the hardest part of rationing is deciding what gets nothing.
That matches what building actually feels like. When I built an LLM reviewer for pull requests, the decisions that moved the needle were not clever instructions — they were choices about what the model never saw. It also explains why sub-agents work: a sub-agent is delegation, a clean context window assigned to one task, returning a short report instead of its raw working notes. Engineering leaders have run this pattern on people forever. The skill of briefing a colleague has never been telling them everything; it is the judgment of deciding what to leave out so that what remains carries weight.
The verdict, and a confession
So, discipline or rebrand — my honest adjudication is that the label was marketing-adjacent and the constraint is physics, and the constraint wins. The most telling detail I found while researching this piece is that the skeptics never wrote their essay. The “it’s just prompt engineering renamed” position lives almost entirely in forum comments and in the rebuttals of proponents quoting it; nobody put their name on the long version. Meanwhile the term kept accumulating institutions: a talk at QCon London this spring, and two weeks ago a dedicated Context Engineering track at the AI Engineer World’s Fair, where — by one attendee’s vendor-blog account — a presenter took the main stage under the job title Context Engineer. Institutionalization proves durability, not truth. The measurements prove the truth.
The confession is that my afternoon of making an agent worse happened months after all of this evidence was published. I could have read my way out of the mistake, and instead I poured the bucket. I spent last June waiting for the term to die of embarrassment, and it turns out the term was waiting for me, because it names exactly what my agent’s decline had already demonstrated. The model was never short of information. It was short of attention — and attention, unlike information, was mine to budget.