Today, if you ask programmers, “Do you use AI at work?”, the answer is essentially a foregone conclusion.

The questions genuinely worth asking are several others.

What are they using it for? How deeply are they using it? How much do they trust it? How widely are the much-discussed Agents actually being used?

Stack Overflow’s 2026 Developer Survey reveals some interesting realities.About 83% of respondents use at least one AI tool at work. Yet only 6.6% said they trust AI for many tasks, including tasks involving important work-related decisions.

This is not evidence that “AI is useless.” Rather, it shows that AI is becoming woven into day-to-day development. However, many people are gradually coming to understand that being able to use it and being able to entrust work to it with confidence are two different things.

This article presents the four areas on the source AI data page—usage, trust and workflows, the Agent ecosystem, and workplace adoption—in clear, easy-to-read English from start to finish. It also includes 22 charts. Rather than translating the official copy word for word, we have summarized it naturally in English while preserving the meaning of the data and survey questions.

First, eight numbers worth remembering

Number	What it indicates
82.8%	Use at least one AI tool at work (about 83%)
65.9%	Use an AI coding assistant or coding Agent
73.0%	Among those who use a coding assistant/Agent, people who use it every day
55.0%	Among those who answered the question about typical AI usage time, use AI in their work for at least two hours every day
62.0%	Positive about using AI at work
48.0%	Trust AI only if its output can be easily verified
9.6%	Use development-related answers generated by AI as-is
53.6%	If AI remembers work-related information, I want to manage that information myself

**First, there is an important caveat about the survey population.** The number of respondents varies by question. Some questions target only people who use AI, while others were asked only of people who have used a specific Agent tool. Also, for questions permitting multiple answers, simply adding the percentages does not result in 100%. The charts below list the sample size for each question.

1. AI has already become an everyday tool for developers

1. Coding assistants are used more commonly than general-purpose chatbots

Among the 17,464 people who answered this question, **65.9% use an AI coding assistant or a coding Agent, while 62.5% use general-purpose AI chat tools.** In other words, AI is no longer confined to a browser-based chat window for asking questions; it has made its way into editors, terminals, code repositories, and development workflows.

However, calling this the “age of Agents” would be an overstatement for now. The proportion of respondents using AI Agents or automated workflows was **26.2%**, so their adoption is not as widespread as that of chat assistants. Furthermore, 17.2% said they do not use AI at work at all.

2. Many people who use coding assistants have already made it a habit

One point here is particularly easy to misunderstand. **The 73% figure does not mean that 73% of all developers use a coding assistant every day.** Rather, it means that 73% of people who use such tools use them every day. The daily usage rates are 65.6% for general-purpose chat tools and 63.1% for Agents/automated workflows.

The frequency of use of internal AI tools is somewhat lower, with 47.7% using them every day. This does not necessarily mean that in-house-developed tools are of lower quality. The scope of their use, access methods, and task frequency may also differ from those of general-purpose tools. At minimum, this shows that whether a tool can be naturally integrated into the workflow is more important than the length of its feature list.

3. AI is not something you just open and use for five minutes a day

Among the 10,926 people who answered the question about daily AI usage time, **30.9% use AI in their work for four or more hours every day, while 24.1% use it for two to four hours every day.Taken together, it is exactly 55%**.

This asks how much of a working day people rely on AI for help, and **it does not mean either the amount of time spent in continuous conversation with AI or that AI single-handedly handled 55% of the work.** Even so, it clearly suggests that, for some developers, AI has changed from a plugin they try occasionally into a tool they use all the time.

4. Overall, the sentiment is positive, but not everyone is wildly enthusiastic

31.6% of respondents were very positive, and 30.4% were somewhat positive, meaning that, in total, 62% were positive. The somewhat negative and very negative groups together accounted for 21.5%, while 15.3% were neutral.

Thus, neither "all developers dislike AI" nor "developers have already accepted AI 100%" accurately reflects reality. A more realistic view is that most people are open to using it, but some are waiting and seeing, evaluating it cautiously, and quite a few clearly dislike it.

2. Increased use does not mean that trust becomes unconditional

5. Conditions for trusting AI—being able to verify the output yourself

If I had to choose just one chart to understand the relationship between developers and AI, this would be the one I would choose.

**48%** said they trust AI only when the results are easy to verify. 16.3% believe AI is reliable for many tasks, but that it should not influence important work-related judgments. Another 16.3% said they would delegate only low-risk tasks to it.Only 6.6% said they truly trust AI across many tasks, including tasks involving important judgments.

This could be called "verifiable trust." AI can present drafts, code, and ideas, but final responsibility and judgment remain with humans.

6. As expected, AI is most commonly used to write and fix code and to write tests

This heat map contains a lot of information, but it is worth taking a closer look. Currently, the most common uses of AI are code generation in familiar domains (69.3%), debugging/troubleshooting/refactoring (63.8%), answering simple technical questions (59.4%), and creating and improving tests (58.1%).

On the other hand, only **19.9% use AI for deployment to production environments, operations, incident response, and similar tasks, while 36%** are not currently using AI and do not plan to use it in the future. AI is also not widely used for writing emails and routine communication documents. **45.9%** said they are not currently using it and do not plan to do so in the future.

This aligns closely with everyday intuition.Having AI write code that can be verified poses less of a psychological burden, whereas letting AI directly touch the production environment is a different matter.

How to read the chart: Responses such as “currently use,” “want to use more in the future,” and “want to reduce usage in the future” were tallied separately and may overlap. Do not interpret each row as a stacked format totaling 100%.

7. Even after receiving an AI-generated response, most people verify it

This is one of the most reassuring figures in the entire report. **76.5%** run the output in a local environment, **63.9%** compare it against the existing codebase, and **53%** check its impact on tests and security. Another 38% consult the official documentation.

By contrast, only **9.6% “use it as is.”** In other words, even if AI has changed how developers write code, engineering discipline in areas such as execution, testing, and review has not automatically been lost.

8. Helpful for complex tasks, but not yet a consistently reliable option

Only **16%** believe AI can handle complex tasks very well. The largest group, at 38.8%, rated it as “okay, but not particularly good.” Another 24% said its performance was low or very low.

For most developers, a realistic approach may be to divide tasks into small, verifiable units and check AI’s suggestions step by step, rather than handing an entire complex requirement to AI and waiting for the results.

9. The AI tools commonly used at work aren’t limited to one

34.4% of people use two AI productseach week, and 40.5% use three to five. Combined, this means that roughly three-quarters of people use two to five types of AI tools each week. However, only 4.8% use six or more types per week.

Be careful not to confuse the “number of models” with the “number of tools.” This question counts the number of products and tools whose AI features people consciously use, not the number of models they switch between within a single product.

3. The agent ecosystem is buzzing, but there is no dominant winner

If you feel that new agent frameworks are appearing practically every day, this chapter should resonate with you. Pay particular attention to the data presented here.Each chart presents data from people who answered that individual question, not market share among developers as a whole.

10. Agent memory layer—competition is still in its early stages

Of the 1,392 people who answered this question, Mem0 ranked first at 20%. It was followed by LangMem (15.2%), Memary (12.6%), and Graphiti and Supermemory (11.6% each).

Since even the leader's share is only 20%, it is clear that memory infrastructure for agents remains fragmented. It will be difficult to reach an early conclusion about which product will become the de facto standard.

11. Agent Orchestration—LangChain Is Still Widely Recognized

**LangChain (30.2%)** ranked first, followed by **OpenAI Agents SDK (23.8%)**. Interestingly, **using the terminal directly (18.9%)** was nearly tied with LangGraph (18.8%).

This result shows that not every agent needs a large-scale framework with multiple layers of abstraction. For many tasks, terminal workflows that are simple, transparent, and easy to debug are also appealing.

12. Observability and Evaluation—Traditional Engineering Tools Also Make Their Mark

For this category of tools, **Sentry (26%)** ranked first, followed by LangSmith at 17.3%, and MLflow and Langfuse at 15.2% each.

Introducing agents is not simply a matter of getting them to “run.” For a team to use them seriously, a series of engineering questions must be answered: Where did the error occur? Why did it give that answer? After changing the prompt, did quality improve or decline? When a failure occurs, how do you track it down? The importance of observability and evaluation is continually increasing.

13. Coding Agents—Claude Code and GitHub Copilot Clearly Lead

Of the 12,255 people who answered the question in this tooling category, **65.5% had used Claude Code over the past year, and 58.7% had used GitHub Copilot**. Next came OpenAI Codex (29.5%) and Cursor (25.5%).

The figures shown here are the percentages of people who have “used them over the past year.” They are neither the proportions of paying subscribers, the daily active user counts, nor ratings of tool quality. Since some people use multiple tools, these percentages cannot be interpreted as mutually exclusive market shares.

14. No-Code Agents—Lovable Takes the Top Spot

**Lovable (38.3%)** took the top spot, followed by Replit (28.3%) and v0 (27.5%); Bolt.new came in at 17.2%.

The value of these tools lies in lowering the barrier to prototyping ideas. However, even after creating a prototype, careful attention must still be paid to data permissions, maintenance, deployment, and security. These figures show usage; they do not mean that “you can confidently launch a production system even if you don’t know how to develop.”

15. Automation workflows—Make, n8n, and Zapier take the top three

In this survey question, Make (38.9%), n8n (32.7%), and **Zapier (18.5%)** took the top three spots. Usage among the remaining tools was more dispersed.

These figures indicate strong demand in the world of agents for “connecting models to real business actions.” Whether a team chooses cloud services, a low-code interface, or an orchestration approach for developers depends on the costs of deployment and maintenance.

4. AI adoption has begun in enterprises, but management practices have not yet kept up

16. AI creates the most value in code implementation and debugging, after all

Combining “useful” and “very useful,” code implementation scored 74.6%, debugging/refactoring 72.6%, documentation 65.7%, and code review 64.5%.

By contrast, only 35% rated communication as “useful” or “very useful,” while 22.6% felt it was not very practical. AI tends to be rated highly for creating, explaining, and correcting structured technical content such as code, whereas its ratings have not risen for communication involving human relationships and context.

17. AI affects not only efficiency but also the experience of work

Thanks to AI, **66.1%** of respondents said their satisfaction with debugging/refactoring work had increased, along with **60.6%** for documentation work and **60.1%** for code implementation.

There is, however, a notable contrast. For code implementation, **19.9%** felt their satisfaction had decreased. AI makes work easier for some people, while others may spend more time reviewing, correcting, and understanding generated results. Becoming “faster” and becoming “more comfortable” are not necessarily the same thing.

18. About two-thirds switched AI tools in the past year

Only **33.7% had never switched tools, and roughly two-thirds had switched at least once.The most common primary reason was the higher quality of the output (25.1%)**, followed by company mandates or recommendations (10.6%), better integration with development environments (7.4%), and lower cost (6.4%).

Importantly, 25.1% is the percentage of everyone who answered this question, not a percentage calculated only among those who switched tools. When choosing a product, developers appear to prioritize the quality of results first, followed by price and other factors.

19. AI “memory”: many people want to manage it themselves

If AI remembers code preferences, project context, and work habits, who should be able to view, edit, and delete that information?

**53.6%** answered “myself.” 14.7% want it to be centrally managed by the company, and 11.1% want it to be managed by the team. Another 14.2% said they do not want AI to have long-term memory in the first place.

This is an important implication for people developing enterprise AI products. More memory functionality is not necessarily better. Rather than “automatically remembering everything,” being transparent, manageable, and deletable may give users greater peace of mind.

20. In many companies, employees are figuring out how to use AI on their own

The most common response, at **29.7%**, was that “AI is essentially optional, and each employee decides how to use it.” 24% said their company has approved tools and clear usage guidelines. Another 15.9% said only some teams are actively adopting it and that usage and management practices are not standardized.

This data cannot be used to directly estimate how many companies worldwide are at a particular stage of governance. This is because the results cover what respondents said about their own organizations. However, it clearly shows that while tool adoption is rapid, rules and collaboration mechanisms have not kept pace.

21. The reasons for consciously not using AI are not limited to disliking new technology

41.3% said they do not consciously avoid AI. Conversely, approximately **58.7%** said they intentionally avoid using AI because of various concerns.

The main reasons were concerns about skills becoming obsolete or jobs being replaced (17.1%), ethical issues (15.2%), privacy and security (13%), and environmental impact (8.2%). Such concerns may not be sufficient reasons to avoid AI entirely, but they clearly affect the scope of use and the extent to which AI is accepted.

Explaining all opposition as simply “not wanting to learn new tools” is clearly too simplistic.

22. Companies are beginning to narrow down their tool choices, but have not reached a unified answer

**32.8%** of respondents said their company had adopted a limited set of company-approved tools. **18.6% said they primarily used only one tool. Combined, 51.4%**, indicating that many teams are beginning to narrow down their tool selections.

Meanwhile, at 17.3% of companies, each team selects its own tools, while 12.5% are still broadly trialing them. Only **4.3%** said that their organization develops its own internal tools or models.

This may be the challenge that many AI projects face next. The issue is not so much a shortage of models as how to create a stable way of working while balancing security, cost, effectiveness, and ease of use.

Finally——After AI becomes widespread, which capabilities will become important?

Looking at these 22 charts, it did not seem to me to mean either that “AI will take over development” or that “AI is not as good as expected.” Rather, it seemed that three changes were happening simultaneously.

**First, AI is changing from a standalone question-and-answer tool into the foundation of the software development process.** Some people are already making extensive use of it for coding, debugging, testing, and documentation. However, the closer it gets to production environments, permissions, security, and critical decisions, the more cautious people become.

**Second, the value of the ability to verify is, if anything, increasing.** AI can produce answers quickly, but you need to know how to execute them, how to check boundary conditions, and how to identify contradictions with the current state of the codebase. The importance of people who understand the system, business operations, and tests is not diminished as a result.

**Third, the Agent market is growing rapidly, but it has not yet taken shape.** While there are clear leaders in coding assistants, there are still various products in areas such as memory, orchestration, observability, and automation. When deploying these at scale within companies, the question is not which model achieves a slightly higher benchmark score, but whether they can build a complete, controllable, and maintainable workflow.

The question for 2026 is no longer “Should developers use AI?” but “How do you strike a balance between mastering its use, deploying it with confidence, and taking responsibility for it?”

(Translated by GPT)

Original link: https://mp.weixin.qq.com/s/BOfSFB3GDN1Z5JVnXtf8Ow?token=603793316&lang=zh_CN&wt.mc_id=MVP_325642