Most of what gets called a trend is a product launch. A few changes this year actually altered the shape of the systems people build.
Long context got cheap enough to be lazy with
When context was expensive, retrieval was mandatory. Now that large windows are affordable, a genuine design choice exists: retrieve precisely, or put more in and let the model sort it out. The second is often faster to build and worse to operate — quality degrades as irrelevant material competes for attention, and cost scales with every call. The interesting systems use long context as headroom for reasoning, not as a substitute for knowing what matters.
Tool access standardised
Two years ago every integration was bespoke. The emergence of a common protocol layer means an agent can discover and call tools it was not built against. The consequence is a shift in where the difficulty sits: connecting is now easy, and the hard problems are permissions, auditability and rollback — the same problems every integration platform has always had.
Evals became infrastructure
Evaluation moved from a research activity to something that runs in CI beside the unit tests. This is the quiet change with the largest effect: it made model upgrades routine. Teams with a regression suite adopt new models in an afternoon; teams without one delay for months because nobody can prove nothing broke.
Small models took the boring majority
The gap between frontier and small models narrowed on exactly the tasks that dominate production traffic: classification, extraction, routing, summarising a short passage. Routing by difficulty is now standard practice, and the frontier model handles a tail that is usually far smaller than teams assume before they measure it.
Agents got narrower and more useful
The general-purpose autonomous agent remains unreliable. What works is narrow: a defined job, a small tool surface, explicit approval on irreversible steps, and a durable record. The systems delivering value this year look less like an artificial colleague and more like a well-instrumented pipeline that happens to reason.
What to take away
The productive shifts were infrastructural — cheaper context, standard connectors, evals in CI, cheaper routing. If you want one thing to adopt from 2026, make it the eval suite. It is what turns every other improvement from a risk into an upgrade.