The state of AI engineering in 2026 is defined by five shifts: agents moved from demos to production with write access, cost became a first-class engineering constraint, multi-model became the default, image generation crossed the adoption threshold, and non-developers started shipping software. These are not predictions — they are the reported behaviors of over 1,000 AI engineers surveyed in a study run by Amplify Partners with Notion and Vercel between May 16 and June 15, 2026 (Amplify Partners, July 2026). The survey had 1,048 respondents, making it one of the largest practitioner-driven datasets on AI engineering to date.
TL;DR — Last verified: July 22, 2026
- 95% of teams use AI agents (roughly double last year); 89% of those agents have write access to data
- 75% of engineers adjust AI usage based on cost; 40% say it regularly limits how ambitiously they build
- 87% use multiple models; open-weight models augment but don't replace closed ones
- Image generation usage doubled from 18% to 36% in one year; audio has the highest intent-to-adopt at 56%
- 97% report net positive organizational impact, but 93% also see negative effects — mainly erosion of deep technical skills
- Over a third of teams now have non-developers shipping features
- Only 4% think the AI engineering stack is working well
Pricing, model versions, and adoption figures change fast — last checked July 22, 2026.
Who Is an AI Engineer in 2026?
AI engineering is no longer a job title — it is a discipline that touches founders, CTOs, product managers, and engineers across company sizes. The survey found that the largest group of respondents (23%) identified as founders or CTOs, followed by AI engineers (19%), fullstack engineers (12%), engineering managers (9%), and product managers (9%) (Notion AI Engineering Survey, 2026).
The experience profile reveals a persistent gap: 51% of engineers with 10+ years of software experience have 3 years or fewer of AI experience. Meanwhile, the newest cohort — engineers with 0-3 years of total experience — have a median of 2 years of AI experience, nearly matching the median AI experience of 10-year software veterans. For new entrants, software engineering experience is AI experience. The share of respondents with less than 1 year of AI experience dropped from roughly 1 in 10 last year to about 1 in 34 this year, meaning the field's leading edge is evening out.
What this means for you: If you're a seasoned engineer feeling behind on AI, you're in the majority. The playing field is leveling fast, and practical experience with AI tools now matters more than years of prior software background.
What Are AI Engineers Actually Building?
Text remains the dominant modality — 96% of respondents are actively building with text. But the more interesting story is in what's changing:
| Modality | Share "working well" | Year-over-year change |
|---|---|---|
| Text | 96% | Baseline |
| Image | 36% | Doubled from 18% |
| Audio | 24% | Up from 17% |
| Video | 25% | — |
Source: Amplify Partners 2026 AI Engineering Report
Image generation saw the biggest jump, driven by improved models like Google's Nano Banana 2 and OpenAI's ChatGPT Images 2.0. What used to produce "cursed hands" is now becoming real work infrastructure.
Audio is the sleeper pick. Among engineers not currently building with audio, 56% say they plan to adopt it — the highest intent-to-adopt ratio of any modality, up from 37% last year. Yet actual adoption growth was modest (17% to 24%), meaning interest is accelerating faster than execution.
What this means for you: If you've been waiting for image generation to get good enough for production work, 2026 is the year it crossed that threshold. Audio is next — expect it to follow the same curve in the next 12-18 months.
How Do Teams Choose AI Models in 2026?
Teams are not picking one model and calling it a day. 87% of respondents actively use multiple models together, and the most popular strategy is routing by task type (44%), followed by running multiple models and comparing outputs (26%), and routing based on cost (11%) (Notion AI Engineering Survey, 2026).
Open vs. closed: not the debate you think
94% of respondents use closed models. 45% use open-weight models. But here's the key finding: over 90% of open-weight users also use closed models. Open weights are augmenting, not replacing. And despite the heated online discourse, the open-vs-closed question is a top-3 consideration for only 5% of respondents when choosing a model.
What actually drives model selection:
- Accuracy / quality — 67%
- Agentic capability (tool calling) — 53%
- Cost — 53%
- Privacy / data — 33%
- Ecosystem / developer experience — 20%
- Reliability / uptime — 20%
Source: Notion AI Engineering Survey, 2026
Reliability ranking low (20%) may seem surprising given public outages, but it likely reflects a threshold expectation: models are reliable enough that the decision moves up the stack to quality, capability, and cost.
The standardization wave
Over 50% of respondents report that their organizations are starting to standardize on fewer AI tools. Some are fully standardizing; others are consolidating some layers while staying flexible on others. The headline: the "Great Standardization" is happening at the platform and tools layer, not the model layer. Teams want fewer tools to manage, but they'll keep multiple models for different tasks.
For a practical approach to multi-model routing, see our guide on which AI model to use for which task.
What this means for you: Don't bet on a single model. Build your architecture to route between models based on task type and cost. Standardize your tools and frameworks, but keep flexibility at the model layer — that's where the field is moving fastest.
Is AI Cost Now a First-Class Engineering Constraint?
Yes. 75% of respondents adjust their AI usage based on cost. 40% say cost regularly shapes how ambitiously they use AI, and another 36% say it sometimes does. Only 24% are rarely or never cost-constrained (Amplify Partners, July 2026).
This is a sharp shift from 2025, when the prevailing attitude was "token maxing is cool" and cost was an afterthought. In 2026, cost shows up as the second most-monitored metric in production (48% of teams actively track it), right behind quality and task success (56%).
| What teams monitor most | Share monitoring |
|---|---|
| Quality / task success | 56% |
| Cost / token usage | 48% |
| User satisfaction | 37% |
| Hallucinations | 37% |
| Tool-call failures | 16% |
| Safety / policy | 15% |
Source: Notion AI Engineering Survey, 2026
Cost is being treated like an SLA — not a budget line item reviewed quarterly, but a real-time metric that shapes product decisions. As model improvements flatten and prices hold steady or rise, this trend will accelerate.
What this means for you: Build cost monitoring into your AI pipeline from day one. Track token usage per request, per user, per workflow. Set budget alerts. Route expensive tasks to cheaper models when quality allows. This is now table stakes, not optimization.
How Prevalent Are AI Agents in Production?
95% of respondents say they're using agents — roughly double last year. But the bigger story is what those agents are doing: among teams using agents, 89% say their agents have write access to data, up from 52% last year (Amplify Partners, July 2026).
Combined, the share of all respondents using write-enabled agents is up more than 3x year-over-year. Agents are no longer just reading, summarizing, and drafting — they're taking actions inside systems.
Permission breakdown among agent-using teams:
| Permission level | Share |
|---|---|
| Read-only | 10% |
| Write with human in the loop | 65% |
| Full autonomy | 25% |
Source: Notion AI Engineering Survey, 2026
This aligns with broader industry data. Gartner predicts that 40% of enterprise applications will feature task-specific AI agents by end of 2026, up from less than 5% in 2025 (Gartner, August 2025). McKinsey's State of AI survey found 62% of organizations experimenting with AI agents, with 23% scaling them in at least one function (McKinsey, November 2025).
Where agents fail
When engineers complain about agents failing, it's usually the thinking, not the plumbing:
- Hallucinations — 61%
- Loses context mid-task — 47%
- Poor reasoning / logic — 37%
Source: Notion AI Engineering Survey, 2026
How teams control agents (and why it's still primitive)
The top guardrail strategies are human-in-the-loop approvals (73%) and access controls / gating permissions (60%). Below that, the results scatter: task decomposition, retrieval, memory, sandboxing. Nobody has settled the control layer for agents.
For a deeper dive on building agents that survive without constant supervision, see our guide on AI agent architecture that survives the 6-month churn. And for keeping agents safe when they have write access, read our 4-layer defense model for AI agent safety.
What this means for you: Agents with write access are now the norm, not the exception. If you're not building guardrails — human approval gates, scoped permissions, audit trails — you're behind. The good news: most teams are also figuring this out, so you're not alone. Start with human-in-the-loop for any action that modifies production data.
What Does the AI Engineering Stack Look Like?
There is no consensus on the biggest challenge in the AI stack. Evaluation leads at 20%, but orchestration/agent logic (18%), inference cost/speed (16%), and security/guardrails (15%) are all within a few points. Only 4% of respondents think the stack is working well (Notion AI Engineering Survey, 2026).
| Stack challenge | Share naming it #1 |
|---|---|
| Evaluation | 20% |
| Orchestration / agents | 18% |
| Inference cost / speed | 16% |
| Security / guardrails | 15% |
| Model quality | 10% |
| Context / retrieval | 9% |
| Observability | 8% |
| "Nothing — it works well" | 4% |
Source: Notion AI Engineering Survey, 2026
Build vs. buy by layer
| Layer | Most common approach | Key insight |
|---|---|---|
| Inference / model serving | Bought (45%) | Most commonly bought layer; few want to build inference infra |
| Prompt management | Built (61%) | Most commonly built — "everyone's prompts are special" |
| Agent orchestration | Built | Product logic stays in-house |
| Fine-tuning | Don't have it | Most haven't adopted; those who bought aren't looking to build |
| RAG / retrieval | Mixed | ~49% use RAG of some sort |
| Evals | Mixed | Still the #1 challenge with no settled solution |
Source: Notion AI Engineering Survey, 2026
The pattern is clear: infrastructure closer to commodity compute (inference, serving) gets bought. Logic closer to the product (prompts, orchestration, RAG) gets built. Fine-tuning is the clearest "not yet" — most teams haven't adopted it at all.
How teams adapt model behavior
- Prompting — 80%
- Tool-use tuning — 59%
- RAG / retrieval — 49%
- Fine-tuning — 26%
- Reinforcement learning — 22%
- Synthetic data — 19%
Source: Notion AI Engineering Survey, 2026
The majority are still just updating prompts — which the report's authors describe as "starting to feel a bit primitive." But it works, and it's the fastest iteration loop available.
What this means for you: Buy your inference infrastructure. Build your prompt management and agent orchestration. Don't worry about fine-tuning until you've exhausted prompting and RAG. And invest in evals — it's the #1 unsolved problem, which means it's the #1 opportunity.
How Is AI Reshaping Engineering Teams?
97% of respondents report a net positive effect from AI on their organization. The top effect isn't just speed — it's cheaper failure. AI made experimentation nearly free: more prototypes, more bets, more ideas tested before committing (Amplify Partners, July 2026).
But it's not free of cost. 93% also report negative downstream effects:
- Erosion of deep technical skills and codebase understanding — the most common complaint, a direct consequence of cheap code generation
- Increased review burden — the same tool that increases experimentation also increases the volume of code to review
- Role blurring — 81% say AI is blurring the line between engineer, product designer, and marketing roles
The most striking organizational shift: over a third of teams now have non-developers shipping features. Mostly smaller, mostly internal — but 17% say non-developers are regularly shipping customer-facing features. Even when non-developers aren't shipping, a third of teams see them building useful prototypes, frontend mocks, and internal tools.
This aligns with data from Augment Code's State of AI-Native Engineering report, which found that 48% of all code is now AI-generated and 63% of engineers are voicing skill-relevance concerns (Augment Code, 2026).
For strategies on managing this shift, see our framework on using AI agents to scale your business and our guide on organizing AI work like an operating system.
What this means for you: AI is democratizing who can ship software, but it's also creating a new kind of technical debt — code that works but that nobody fully understands. Invest in code review processes that can handle higher volume. Consider decoupling your agent architecture so that model changes don't require rewrites.
What Do AI Engineers Predict for the Next 5 Years?
The survey asked respondents to place bets on the future:
- 76% say AI boosted their job satisfaction
- 59% fear today's AI code creates long-term liabilities
- Only 33% call software engineering a solved problem
- 67% expect a leading lab will declare AGI in the next 5 years (note: the question asked about the press release, not the achievement)
- Only 9% bet on Transformers being state-of-the-art in 5 years — most are unsure
- 36% yes / 38% no on whether there will be more AI compute in space or on land — the most divisive question in the survey
Source: Amplify Partners, July 2026
The tension between happiness (76%) and fear (59%) captures the moment: engineers are faster and more productive, but they know they're accumulating maintenance debt. "Happier, faster, but embracing the maintenance burden" is the TL;DR.
How to Prepare Your Team for AI Engineering in 2026
Based on the survey data and primary sources, here's a concrete action plan:
- Adopt multi-model routing. 87% of teams already use multiple models. Route by task type, fall back by cost. Don't lock into a single provider.
- Build cost monitoring from day one. 48% of teams now monitor cost in production. Track token usage per request, set budget alerts, route to cheaper models when quality allows.
- Give agents write access — with guardrails. 89% of agent-using teams already do. But pair it with human-in-the-loop approvals (73% of teams do) and scoped permissions (60%).
- Invest in evals. It's the #1 unsolved challenge (20% name it their top problem). The "vibe review" is still the most common method, which means good evals are a competitive advantage.
- Prepare for role blurring. Over a third of teams have non-developers shipping. Create lightweight processes for non-engineers to contribute code safely — sandboxed environments, review gates, clear ownership.
- Watch image and audio modalities. Image generation crossed the adoption threshold (36%, up from 18%). Audio has 56% intent-to-adopt. If your product could benefit from either, now is the time to prototype.
- Standardize tools, not models. Over 50% of orgs are consolidating their AI tooling. Pick your frameworks, your eval tooling, your monitoring — but keep the model layer flexible.
FAQ
Q: What percentage of teams use AI agents in 2026?
A: 95% of AI engineering teams report using agents, roughly double the previous year. Among those teams, 89% give agents write access to data, up from 52% the year before. The overall share of respondents using write-enabled agents is up more than 3x year-over-year, according to the Amplify Partners 2026 AI Engineering Report.
Q: How many AI engineers use multiple models?
A: 87% of respondents actively use multiple models together. The most popular routing strategy is by task type (44%), followed by running multiple models and comparing outputs (26%), and routing based on cost (11%). Only 13% use a single model.
Q: Is cost a major concern for AI engineering teams?
A: Yes. 75% of respondents adjust their AI usage based on cost, and 40% say cost regularly shapes how ambitiously they build. Cost is the second most-monitored metric in production (48% of teams), right behind quality and task success (56%).
Q: Are open-source AI models replacing closed models?
A: No. 45% of respondents use open-weight models, but over 90% of those also use closed models. Open weights are augmenting, not replacing closed models. The open-vs-closed debate is a top-3 consideration for only 5% of respondents — quality, agentic capability, and cost matter more.
Q: What is the biggest challenge in the AI engineering stack?
A: Evaluation leads at 20%, but orchestration/agent logic (18%), inference cost (16%), and security (15%) are close behind. Only 4% of respondents think the stack is working well. The scattered responses indicate the field is still maturing — every layer has unsolved problems.
Q: How is AI affecting engineering team roles?
A: 97% report a net positive organizational impact — mainly faster experimentation and cheaper failure. But 93% also see negative effects, primarily erosion of deep technical skills. Over a third of teams now have non-developers shipping features, and 17% say non-developers regularly ship customer-facing code.
Q: Will AI labs declare AGI within 5 years?
A: 67% of survey respondents expect a leading lab to declare AGI in the next 5 years. However, the survey specifically asked about the press release, not the achievement — meaning most engineers believe the announcement will happen regardless of whether true AGI is reached. Only 9% believe Transformers will remain the state-of-the-art architecture in 5 years.
Discussion
0 comments