# State of AI engineering 2026: agents, evals, costs and regulation

URL: https://sdclabs.com/state-of-ai-engineering-2026/

> AI engineering in 2026 is about reliability in production: 88% of organisations use AI, but agent deployment is in the single digits in most functions and quality is the top barrier.

# The state of AI engineering in 2026

[Mykhailo Poliarush](https://sdclabs.com/author/poliarush/) 2026-11-176 min read

In 2026, AI engineering is no longer about getting a model to work. It is about making AI systems reliable in production. Models now complete most real coding and computer tasks in public benchmarks, and 88% of organisations use AI in at least one business function, according to the [Stanford AI Index 2026](https://hai.stanford.edu/assets/files/ai_index_report_2026.pdf) . Yet agents are deployed in the single digits in nearly every function, and quality is the top barrier to production. This report collects verified data on where AI engineering stands in 2026, with every figure linked to its primary source.

## How good are AI models at real software work in 2026?

Capability rose sharply in a year. On SWE-bench Verified, which measures fixing real GitHub issues, the top score was about 76.8% in February 2026 (Claude 4.5 Opus), with several models clustered in the low-to-mid 70s, according to the AI Index 2026. On OSWorld, which tests real computer tasks, the best agents went from roughly 12% to 66.3%, within six percentage points of computer science students at about 72%.

OSWorld: share of real computer tasks completed

**Best AI agent, earlier**

12%

**Best AI agent, 2026 (Claude Opus 4.5)**

66.3%

**Computer science students**

72%

Source: [Stanford HAI, AI Index Report 2026 (pp. 72, 113)](https://hai.stanford.edu/assets/files/ai_index_report_2026.pdf)

Benchmarks need care. In February 2026, OpenAI [stopped reporting SWE-bench Verified](https://openai.com/index/why-we-no-longer-evaluate-swe-bench-verified/) after an audit found material issues in 59.4% of 138 problems it checked. Public scores show direction; your own evaluations on your own code show fit.

## Are AI agents in production, or still in pilots?

Both, depending on who you ask. In the AI Index 2026, which draws on McKinsey survey data, 88% of organisations used AI in at least one function in 2025, up from 78% in 2024, but AI agent deployment was in the single digits across nearly all business functions. Among teams that build with agents, adoption is higher: in [LangChain's State of Agent Engineering](https://langchain.com/state-of-agent-engineering) survey of 1,340 practitioners, 57.3% have agents in production, and 32% name quality as the main barrier.

Developers keep agents supervised. In the [Stack Overflow Developer Survey 2025](https://survey.stackoverflow.co/2025/ai) , 14.1% used AI agents at work daily and 37.9% did not use them and had no plans to. A [May 2026 Stack Overflow pulse survey](https://stackoverflow.blog/2026/05/27/agents-on-a-leash-agentic-ai-remains-mostly-monitored-at-work) of about 1,100 developers found that 63% rarely or never let agents run on autopilot, and 60% block agents from making unapproved system changes.

## What does an AI engineering stack need besides a model?

Evaluation and observability have become standard practice. In the LangChain survey, 89% of teams have observability for their agents, 52.4% run offline evaluations and 37.3% run online evaluations in production. Teams that skip evals are flying blind on the very problem, quality, that blocks most deployments.

How teams run AI agents (LangChain survey, 1,340 responses)

**Have observability for agents**

89%

**Have agents in production**

57.3%

**Run offline evaluations**

52.4%

**Run online evaluations**

37.3%

**Name quality as the top barrier**

32%

Source: [LangChain, State of Agent Engineering (fieldwork Nov to Dec 2025)](https://langchain.com/state-of-agent-engineering)

Integration standards are settling too. The Model Context Protocol, which connects models to tools and data, had more than 10,000 active public servers and more than 97 million monthly SDK downloads when Anthropic [donated it to the Agentic AI Foundation](https://www.anthropic.com/news/donating-the-model-context-protocol-and-establishing-of-the-agentic-ai-foundation) under the Linux Foundation in December 2025 (Anthropic's own figures).

Cost keeps falling. [Epoch AI](https://epoch.ai/data-insights/llm-inference-price-trends) measured LLM inference prices dropping between 9x and 900x per year depending on the performance level, and for GPT-4-level performance on PhD-level science questions, about 40x per year. Epoch notes that the fastest declines may not continue.

## Do developers trust what AI writes?

Not by default. In Stack Overflow 2025, 84% of developers used or planned to use AI tools, but 46% distrusted the accuracy of the output against 33% who trusted it. The most cited frustration, at 66%, was AI answers that are "almost right, but not quite". In the [DORA 2025 report](https://blog.google/innovation-and-ai/technology/developers-tools/dora-report-2025/) , 90% of software professionals use AI and more than 80% say it improved their productivity, yet 30% trust AI-generated code only a little or not at all.

Safety data points the same way. The AI Incident Database recorded 362 AI incidents in 2025, up from 233 in 2024, as reported in the AI Index 2026, and the average Foundation Model Transparency Index score fell from 58 in 2024 to 40 in 2025.

## Does AI actually make engineers faster?

The evidence is mixed. In a [2025 randomized trial by METR](https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/) , 16 experienced open-source developers took 19% longer on 246 real tasks when they used AI, while believing they had been 20% faster. METR's [February 2026 follow-up](https://metr.org/blog/2026-02-24-uplift-update/) estimated an 18% speed-up for returning developers and 4% for new recruits, but both confidence intervals include zero and METR calls it very weak evidence.

Hiring data shows that the work is changing anyway. The AI Index 2026 reports that mentions of generative AI skills in US AI job postings grew 111% from 2024 to 2025, and that by September 2025 employment of US software developers aged 22 to 25 had fallen close to 20% from its 2022 peak. That last figure is a correlation, not proof that AI caused it.

## Where is enterprise AI money going?

Spending is concentrating on a few model providers. [Menlo Ventures' 2026 State of Enterprise AI](https://menlovc.com/perspective/2026-the-state-of-enterprise-ai/) reports Anthropic's share of enterprise LLM API spend at 55% in 2026, up from 40% in its [2025 report](https://menlovc.com/perspective/2025-the-state-of-generative-ai-in-the-enterprise/) , with OpenAI at 31% and Google at 9%. Menlo Ventures is an investor in Anthropic, and its surveys cover a few hundred respondents, so treat these as one investor's estimate.

Share of enterprise LLM API spend by model provider (Menlo Ventures)

|   | 2025 | 2026 |
|---|---|---|
| Anthropic | 40% | 55% |
| OpenAI | 27% | 31% |
| Google | 21% | 9% |

Source: [Menlo Ventures, State of Enterprise AI 2026 and 2025 (Menlo is an investor in Anthropic)](https://menlovc.com/perspective/2026-the-state-of-enterprise-ai/)

The same 2025 report found that 76% of AI use cases were bought rather than built, and that only 16% of enterprise deployments and 27% of startup deployments were true agents. Most companies buy AI products and wrap them into their own systems, which makes integration, testing and governance the engineering work that matters.

## What does EU regulation require, and when?

The [EU AI Act](https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai) applies in stages. Prohibited practices have applied since 2 February 2025 and obligations for general-purpose AI models since 2 August 2025. The AI Omnibus amendment, in force since 27 July 2026, moved the high-risk obligations for Annex III use cases to 2 December 2027 and for AI in regulated products (Annex I) to 2 August 2028.

EU AI Act: when obligations apply

|   | Date |
|---|---|
| AI Act enters into force | 2024-08-01 |
| Prohibited AI practices apply | 2025-02-02 |
| General-purpose AI model obligations apply | 2025-08-02 |
| AI Omnibus amendment takes effect | 2026-07-27 |
| High-risk obligations, Annex III use cases | 2027-12-02 |
| High-risk obligations, Annex I (AI in regulated products) | 2028-08-02 |

Source: [European Commission, AI Act](https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai)

For teams building on AI in or for the EU, the practical steps are the same as good engineering practice: document what the system does, test it, log it and keep a human able to intervene.

## What should engineering leaders do with this?

Treat AI engineering as engineering. Pick models with your own evals, not leaderboards; put observability and evaluation in place before scale; keep agents supervised where changes are hard to undo; and verify AI-generated code with the same automated tests and reviews as any other code. That is how SDCLabs works: AI agents do the drafting, senior engineers stay in control, and test automation from the founders of Codeception and CodeceptJS checks every change. See our [AI engineering services](https://sdclabs.com/services/ai-engineering/) .

Sources were checked on 2026-11-17. Survey figures describe their respondents; vendor and investor figures are labelled as such.

## Frequently asked questions

### What is AI engineering?

AI engineering is the practice of building software products with and on top of AI models: choosing models, connecting them to data and tools, evaluating their output, monitoring them in production and keeping humans in control of what they change.

### Are AI agents ready for production in 2026?

Some are. 57.3% of teams in LangChain's survey have agents in production, but agent deployment is in the single digits in nearly every business function in the AI Index 2026, and quality is the most cited barrier. Supervision, evals and observability are what make the difference.

### When do the EU AI Act high-risk rules apply?

After the AI Omnibus amendment, high-risk obligations apply from 2 December 2027 for Annex III use cases and from 2 August 2028 for AI in regulated products under Annex I. Prohibitions and general-purpose AI model obligations already apply.

[#ai](https://sdclabs.com/tag/ai/) [#ai-agents](https://sdclabs.com/tag/ai-agents/) [#development](https://sdclabs.com/tag/development/)

[Mykhailo PoliarushFounder and CEO](https://sdclabs.com/author/poliarush/)

## Related posts

[
### AI in the software development lifecycle: 7 trends backed by data

AI is now used at every stage of the SDLC, but unevenly: 90% of technology professionals use it, while 30% report little or no trust in AI-generated code.

AI engineering · 2026-11-17

](https://sdclabs.com/ai-sdlc-trends-2026/)

## Get an estimate for your product.

[Get an estimate](https://sdclabs.com/contact/) [Book a call](https://sdclabs.com/contact/#call)
