Amplify Partners' annual State of AI Engineering survey drew 1,048 respondents this year, run with Notion and Vercel. The headline shifts: image generation adoption doubled, cost became a first-class engineering constraint, and write-enabled agents more than tripled year over year while guardrails stayed primitive.
Bar opened by noting how fast the field moves — "the field changes as you make the slides" — citing recent Frontier releases treated like national security events and reports of Meta exploring selling AI compute. That volatility is exactly why Amplify runs the survey each year: to step back and understand what AI engineers are actually doing. For the first time, the survey was run in partnership with Notion and Vercel.
This year drew 1,048 respondents. Bar stressed that AI engineering is "more of a discipline than a job title" — the sample spans founders, CTOs, engineers, and product people across company sizes. For the third year running, respondents skew senior in software but newer to AI: over half of those with 10+ years of software experience have three years or less of AI experience, and the newest engineers have never known software without it.
Text dominates the modalities people build with, but Bar tracks the "intent to adopt" ratio — of those not using a modality, how many plan to. Audio leads, with 56% of non-users planning to adopt (up from 37% last year). The biggest year-over-year jump, however, was image generation: the share using generative AI for images and feeling good about it doubled from 18% to 36%, which she tied to launches like Nano Banana, Nano Banana 2, and ChatGPT Images 2.0.
On models, 94% use closed models and 45% use openweight models, but openweight is augmentation, not replacement — over 90% of openweight users also run closed models, and 87% of teams use more than one model. What drives model choice is quality first, then agentic capabilities (tool calling) tied with cost. Despite all the open-vs-closed airtime, that was a top-three factor for only 5%, and reliability ranked low — likely because it has become a threshold requirement met by most models.
"Infinite intelligence still comes with a usage-based bill." Bar framed cost as a first-class engineering constraint: 40% of respondents say cost regularly shapes how ambitiously they use AI and another 36% say it sometimes does — roughly three in four adjusting usage based on cost. It shows up in monitoring too, where cost and token usage is the #2 thing watched, "right under quality itself."
Agents are the biggest line item. This year 95% report using agents (roughly double last year), and among agent builders, those whose agents can write data jumped from 52% to 89% — so overall write-enabled agent usage more than tripled. "Agents are no longer reading, summarizing, drafting. They're taking actions inside of systems." Yet control remains blunt: the top two mechanisms are human-in-the-loop approvals and gating permissions, and two-thirds say hallucination or losing context mid-task is their biggest frustration.
Every year the #1 stack challenge is eval — still true, but by a thinning margin, with 96% reporting some stack problem and "vibe review" the top evaluation method. On build-vs-buy across eight layers, inference and model serving is bought the most, while prompt management is built in-house by 61% ("everyone's prompts are special"), and fine-tuning is mostly "not yet."
For teams, 97% report a net-positive organizational effect — the top benefit being cheaper failure and more experimentation rather than raw speed — but over 9 in 10 also feel negative downstream effects like erosion of deep technical skills. 81% say AI is blurring engineering with product, design, and marketing; over a third of teams have non-developers shipping features, and 17% say non-developers regularly ship customer-facing ones. On five-year bets: 76% say AI boosted job satisfaction, 59% fear today's AI code creates long-term liabilities, only a third call software engineering solved, 67% expect a lab to declare AGI within five years, and just 9% bet Transformers stay state-of-the-art — while the most divisive question was whether more AI compute will be in space or on land (36% yes, 38% no).
We had 1,048 respondents this year, which is a lot of AI engineers.
Cost is now a first class engineering constraint.
Agents are no longer reading, summarizing, drafting. They're taking actions inside of systems.
The most divisive question in the survey is about outer space.
| Time | Topic |
|---|---|
| 00:49 | Why run the survey every year; partnering with Notion and Vercel |
| 02:21 | 1,048 respondents — a discipline, not a job title |
| 03:23 | Modalities: text dominates, audio's intent to adopt, image gen doubles |
| 05:28 | Models: closed vs openweight, quality drives choice |
| 08:34 | Cost becomes a first-class engineering constraint |
| 09:35 | Agents get write access; blunt control mechanisms |
| 12:11 | The stack: eval, build vs buy |
| 14:17 | Effect on teams and rapid-fire five-year bets |