State of Development Report 2026

What are they using AI for?

49.1% say agents are “in production” or “core to how they ship”

The AI-using cohort has traveled deeper into agent territory than we thought.

Frequent AI agent use jumped 70.8% in the last year (daily or more)

For comparison, it took Agile 11-15 years to reach mainstream. Here, we’ve seen AI agent usage increase from 47.3% saying “daily or more” to 80.8%.

The successful were 1.5x more likely to use agents daily or more (85.5% vs 55.8%) and 2.3x more likely to use agents continuously (23.9% vs 10.4%).

The median respondent is using 5.0 agents

Though on average, they use 10.7. That 2.1x difference between the mean and median indicates that there are a few very large outliers — individuals or teams running legions of agents, such as those who replied “256.”

The successful run 1.3x more agents on average (11.1 vs 8.8).

The primary use cases are writing code, testing code, analyzing, generating ideas, and debugging

After that, there’s a modest dropoff to the bulk of lesser use cases — though none was used by less than 14.8% of all respondents.

The successful are much more likely to use AI agents for technical design, security, and support. They also use AI agents for more use cases.

The successful are more likely to use AI for:

• 2.7x for technical design (28.0% vs 10.5%)
• 2.6x for security (33.5% vs 12.8%)
• 2.2x for customer support (28.0% vs 12.8%)
• 1.6x for writing code (66.7% vs 40.7%)
• 1.5x for researching knowledge bases (45.9% vs 31.4%)

Do teams want to use AI differently? Not really.

We asked engineering teams what they want to be using AI agents for, to compare, and were surprised to find little difference. Is this because the future is unclear? Or that roles and use are changing too quickly to see ahead? Or that they simply already know what agents are good for? (Write-in responses suggest the final option is likely.)

If we zoom in, the successful are somewhat more certain that they’re using AI agents the way they want — there’s less of a gap between their present usage and desired usage.

The top-used AI tools: #1 ChatGPT or API, #2 Microsoft Copilot, #3 Gemini, and #4 Claude

OpenAI has maintained its first-mover advantage and first place. It also has more offerings. If we combine the model offerings from OpenAI and include services by others like Microsoft’s Copilot that run on OpenAI’s tools, it accounts for most of the list.

CoPilot and Gemini’s success are likely due to distribution — those are widely embedded in work tools, and defaults can be powerful. Claude is a close fourth, just above ChatGPT Codex.

The average team is using 3.5 of these tools.

The successful are using 1.3x more tools than everyone else (3.6 tools vs 2.7 tools), suggesting that when you get good at orchestration, you’re either able to use more tools — or have to.

The successful are more likely to use:

• AWS Bedrock (23.1% vs 8.1%)
• Azure OpenAI Service (30.3% vs 18.6%)
• ChatGPT Codex (34.4% vs 24.4%)
• Claude (41.2% vs 31.4%)
• GitHub Copilot (32.7% vs 15.1%)
• Google Gemini or Vertex AI (44.2% vs 29.1%)
• LangChain or LangGraph (7.3% vs 2.3%)

Temporal users are 3.1x more likely to use open source models (37.0% vs 12.4%). They also use more tools on average (5.0 versus 3.5).

Agent frameworks: Again, OpenAI products were the most commonly selected

Compared to last year, Microsoft’s agent framework made gains, as did AWS Strands and CrewAI.

The successful are 1.7x more likely to be building agents on OpenAI Agents SDK (51.1% vs 30.2%) and are 1.9x more likely to use Temporal (9.0% vs 4.7%).

Orchestration: We’re seeing falling interest in AWS Step Functions, LangGraph, and Airflow

Companies are only modestly looking to change how they orchestrate their agents, with declining interest in AWS Step Functions and LangGraph, and rising interest in Camunda, Inngest, and Prefect. Temporal, Dagster, and Orkes held roughly constant.

The successful show less of a desire to change how they orchestrate than everyone else. If it’s not broken, don’t fix it.

What effect have AI agents had on you and your organization?

91.1% say AI agents have ‘improved’ or ‘revolutionized’ their productivity

Nearly everyone agrees that agents make developers more efficient. Only 4.0% said they had no impact and just 1.8% said agents had worsened their productivity.

51.3% go from AI prototype to production-ready code in hours or faster

77.1% say they get there in days or faster. This is a striking finding given that 33.4% of our dataset were respondents from companies with 1,001+ employees; 14.6% of companies had 5,001+ employees. So this isn’t just a function of startups.

If we look at just that cohort of respondents from companies with 5,001+ employees, they fall into two categories: The very fast and the very slow. Most use agents to deploy at startup speeds, while one-fifth take “months” (18.5% replied that way, versus 5.5% of all others).

The successful are only 1.2x faster

A number of engineers told us that their biggest bottleneck is just plain old time to think. That shows up in several of the results where the successful aren’t substantially faster (78.9% daily or faster vs 67.4%). Yet their outcomes are better. AI speeds everyone up equally. Some make better use of it.

85.5% trust the outputs at least somewhat

Trust is high. One quarter (24.7%) completely trust the outputs. A majority (60.8%) at least somewhat trust. Only 4.0% distrust. These levels of confidence are surprising given all the public discourse around code slop and hallucinations.

The successful are far more trusting: They are 6.1x as likely to ‘completely trust’ (28.4% vs 4.7%) and 1.6x as likely to ‘somewhat trust’ (90.8% vs 57.0%).

Temporal users are 1.1x as likely to say they somewhat or completely trust their agents (95.7% vs 84.7%).

84.5% believe they are better at using agents than their competitors

A great majority of teams we polled think they’re in the 85th percentile of agent excellence which is, statistically, unlikely. (Possible, but unlikely.) We choose to interpret this favorably: There is a lot of misinformation out there about agent use. Most teams aren’t informed about what others are actually up to, hence the need for this report.

No surprise, though, some of this exuberance is driven by leadership: CEOs, management, VPs, and directors were 1.4x more likely to say their organization is ‘very successful’ (43.7% vs 32.4%).

Temporal users say they are 1.1x as likely to say they are successful at using agents than peers (91.3% vs 83.8%).

Leadership is 35.1% more likely to say their organization is ‘very successful’ compared to the engineers.

41.1% encounter issues with AI agents daily or more

16.4% say they encounter issues hourly or more, which sounds punishing. Surprisingly, the number of agents teams deploy and the frequency of issues are not correlated. Teams running dozens of agents do not necessarily encounter more issues than those running five. Though, a caveat: Each team may define issues differently, issues can range in severity, and those who pay closer attention will notice more of the granular issues.

Successful teams are just as likely as others to experience issues “daily or more.” But expand that and you see that the successful are 1.7x more likely to experience issues “continuously” (more frequently) and everyone else is 1.3x more likely to experience issues daily (less frequently).

YouTube is the #1 place teams go for help troubleshooting agents

It is interesting that YouTube, a video site where viewers cannot as easily copy the code they see onscreen, beats GitHub and Stack Overflow, where they can. (It is possible to copy code from YouTube with Gemini, but less convenient. It sometimes pulls only fragments). This suggests that AI agent developers would rather watch someone walk through the results onscreen than copy a pattern. Maybe this means the temporal aspect of running many agents makes it more important to understand the workflow than the exact code, which they can easily generate.

In future research, we are curious to know whether this suggests a trend toward isolation: Engineers will look at videos, Discords, chat groups, academic sources, GitHub, and X (Twitter) all before they talk to someone at their own organization. This sense of collaborative distance shows up in the written responses.

Also: 1.8% of people think they’re in the smartest 1%. Talk about accuracy.

Engineers will look at videos, Discords, chat groups, academic sources, GitHub, and X (Twitter) before they talk to someone at their own organization.

The successful are more resourceful troubleshooters: They search 4.1 places, where the others only search 3.5 places.

The successful are also:

• 2x as likely to turn to an AI tool for help (40.4% vs 19.8%)
• 1.9x as likely to turn to Twitter (31.6% vs 16.3%)
• More likely to turn to peers, LinkedIn, Stack Overflow, Reddit

92.3% have tried building an app they’d previously have bought

25.6% said they succeeded in building an app they’d otherwise have paid for, and that it had a big impact. This suggests the so-called “SaaSpocalypse” is real — teams are finding the tools to build new tools and aren’t coming back as frequently for resupply.

The successful are 8.5x more likely to have rebuilt something and said it had a big impact (29.7% vs 3.5%). Everyone else is 2.9x as likely to say they didn’t try or it didn’t work (17.0% vs 6.0%).

Temporal users are 1.6x more likely to say that they’re rebuilding rather than buying and it had a big impact (39.1% vs 24.4%).

For future research: How will devops teams adapt when there are far more custom/vibe-coded tools that don’t necessarily interoperate? Does all this self-generated tooling make it harder for developers to switch jobs? Does it mean they’ll require longer to onboard at the new one?

What’s holding teams back from using AI agents more?

Top blockers to using AI agents more: #1 tracking state, #2 debugging, #3 managing costs

In our survey, 35.7% said the most common blocker to using more AI agents is tracking state, followed by debugging (agents or otherwise) and managing costs (tokens or compute).

Successful teams had more responses per respondent, meaning they face a wider variety of issues — or at least are better aware of them. They were somewhat more likely to say, “Tracking,” “retries,” “unrelated business logic,” and “infrastructure,” and somewhat less likely to say, “orchestrating,” “isolating,” or “managing costs.”

Where are the truly autonomous agents? 39.5% cite security concerns

The greatest blocker to launching self-managing agents is the same thing that has always blocked teams from doing more outsourcing: Engineers need certainty that their delegates have done their job correctly. The only way to know — beyond time-consuming testing and mathematical proofs — is to manually check.

Thus, even while 85.5% of respondents trust the agents’ outputs at least somewhat, there is a difference between trusting enough to proceed and trusting enough to put your name on the commit.

Successful teams worry somewhat less about the unknowns, such as lack of reliability or invented results, unintended consequences, and unpredictable outputs.

Temporal users are 2.2x as worried about AGI (21.7% vs 10.4%). They are 1.4x less likely to be blocked by poor integrations with existing systems (15.2% vs 22.6%) and 2.5x more worried about the climate impacts of AI agents (19.6% vs 7.9%).

Asked to elaborate, respondents volunteer more skepticism and uncertainty. They acknowledge more risks and drawbacks.

79.8% say token and compute costs limit their progress

Companies are realizing with chagrin that in some cases, tokens can be quite expensive. The successful are 1.3x more likely to say cost is at least somewhat of a factor (83.1% vs 65.7%). Everyone else is more likely to say they’re unsure (22.1% vs 7.3%).

How are they feeling about it?

44.6% of respondents were less stressed than one year ago

We expected panic. Instead, we found a spacey calm. Even at the frontier of progress, life goes on. There’s gravity. People need to eat. Respondents are more likely to say they are less stressed compared to a year ago.

Teams that are successful are 2.2x more likely to say they are feeling less stressed than one year ago (48.7% vs 22.1%), and are less likely to say they are more stressed (30.4% vs 44.2%). This is because everyone else is so likely to say that nothing has changed for them year-over-year. Whereas for the successful, things are changing one way or the other.

The more stressed the team, the more issues they encounter: 52.2% of those who are more stressed encounter agent issues daily, versus 32.4% of those who are less stressed.

Many say agents do not factor into their stress

At the frontier, software engineering is difficult and agents aren’t the biggest factor. Those in technology can sometimes forget that the new thing isn’t everything, and that managing their time, bureaucracy, organizational disorder, tooling, meetings, and testing are still the bigger issues. And perhaps will always be.

People are optimistic about a future with more agents, but also worried

77.5% are more optimistic about the future of their role than a year ago

The successful are 8.4x more likely to say they are “much more optimistic” (39.1% vs 4.7%). Though both groups are equally likely to say “slightly optimistic.”

The successful are 6.4x as likely to say they are much more optimistic about the future for their peers, too (37.2% vs 5.8%).

Temporal users are 1.2x more likely to say they are optimistic about the future of their role (89.1% vs 76.4%).

Despite the fear for others’ jobs, only 26.4% said their companies were ‘stopping’ or ‘slowing’ their hiring

Everyone thinks the prospects for hiring are worsening. But their own companies tell a somewhat less pessimistic story: Most are still hiring, and many are prioritizing AI agent experience. The spaceship, it seems, is well-insulated.

The successful are more likely to be hiring, and hiring evenly: They are equally likely to be hiring for senior as for junior roles. They are also 1.8x more likely to be hiring for “AI agent” experience (47.9% vs 26.7%).

Everyone else, on the other hand, is 1.7x as likely to be hiring for junior roles as they are senior roles, or are waiting to see how things play out: They are 2.7x more likely to say AI agents have had “no impact” on their hiring plans (20.9% vs 7.7%)

29.6% think agents are getting too much societal attention

Everybody — 98.5% of respondents — had an opinion on this. The successful are more likely to feel the societal attention is correctly calibrated, and less likely to have “no opinion.”

A large number of developers think AI agents will have a positive ecological impact

Will AI usher in an era of scientific discoveries that solve climate change? Or will it just make things worse in the fruitless pursuit? Developers lean toward the former.

Though, recall that 8.8% cited ‘societal or climate’ worries as a reason for not using AI agents more. Those who felt that agents would have a negative climate impact tended to fall into the everyone else cohort, which is not commentary on our part, just an output from the survey.

Looking 5-10 years out, teams think AI reliance is inevitable

In a world with more AI agents, they largely agree that engineers will act more and more like “natural language orchestrators” over fleets of agents that handle code generation, integration, and maintenance.

Themes among these responses:

• It will invite in human creativity
• Code will automatically integrate
• Programming will be instantaneous
• The “Developer” and “AI trainer’ distinction will be blurry
• There will be many 1-person “teams”
• Code-gen will become far more inclusive to the non-technical
• There will be fewer jobs, including for engineers

Their advice to us in the present? Many say “chill out”

Answers of course range, and include responses like “panic,” “boycott, “fearmongering,” and “worry,” but the most common responses were (semantically): relax, adapt, learn, chill, calm, focus, use, embrace, learning, and prioritize.

What does the agentic near future hold?

The good and the bad reactions net out to a whole lot of positivity. When asked to reflect, engineers do express worry, but it pales next to their optimism about their roles and personal futures.

What can teams in the present take away from the words of those at the frontier? That successful teams are less worried and less stressed even while facing the same rate of errors. Their success compounds as they use more tools, search more widely for help troubleshooting, and build more of their own developer applications. This may be a look at what some are calling “the great separation” between avid adopters and the rest, as results beget more, automatically, ad infinitum.

Whether causation or correlation, successful teams are:

• Using 1.2x more tools
• More likely to trust AI agent outputs (6.1x)
• Not necessarily faster (1.2x) — just more efficient
• More worried about compute costs (1.3x)
• Less worried about their job prospects (8.4x)
• Less worried about the job prospects of others (6.4x)
• Less stressed overall (2.2x)

Full methodology

Between April 29 to May 25, 2026 we commissioned the survey provider Qualtrics to solicit responses from 650 respondents currently using AI agents following strict criteria around role and location. We removed all responses we considered low quality and were left with 554 respondents.

• Good spread of company sizes, concentrated in mid-sized firms (251–1,000 the largest group)
• Decent spread of roles; Engineer/AI engineer most common (25.6%)
• Industries biased toward software (37.5%)
• Experience skews mid-to-senior: 32.9% have 6–10 years, 48.4% have 11+ years
• Heavily US (two-thirds) with the remaining third UK/EMEA