AI agents now rank among the top tools developers use for writing and testing code, but manual checks still gate what reaches production.
The State of Development 2026 report from Temporal puts numbers behind that pattern, drawing on a survey of 554 engineers and technical leaders who currently use AI agents. Its opening finding sets the scale of adoption: 49.1 percent of respondents rate agents as “in production” or “core to how they ship.”
Temporal’s report splits most findings into two groups: a “successful” cohort that rates its own use of agents as successful relative to peers, and everyone else. “Daily or more” usage climbed from 47.3 percent a year ago to 80.8 percent today, with continuous use rising alongside it.
Successful teams run agents daily or more at 85.5 percent, against 55.8 percent for everyone else, a gap of 1.5 times. They run agents continuously at 2.3 times the rate of everyone else, 23.9 percent against 10.4 percent. The median respondent operates five agents, though the mean of 10.7 points to outliers, including one who put the figure at 256. Successful teams run 1.3 times more agents on average than everyone else, at 11.1 against 8.8.
Testing and debugging dominate agent workloads
Writing code, testing code, analysing, generating ideas and debugging make up the primary use cases for AI agents. Every other use case still cleared 14.8 percent of respondents, so the drop-off toward less common tasks stays modest.
Successful teams pull further ahead on specific tasks. They are 2.7 times more likely to use agents for technical design, at 28.0 percent against 10.5 percent, and 2.6 times more likely for security work, at 33.5 percent against 12.8 percent.
Customer support shows a 2.2 times gap, at 28.0 percent against 12.8 percent. Writing code and researching knowledge bases show smaller gaps, 1.6 times at 66.7 percent against 40.7 percent and 1.5 times at 45.9 percent against 31.4 percent.
Desired use looks close to actual use in the report’s side-by-side comparison, with little space between what teams do with agents today, and what they say they want to do with them. The report attributes this to teams already knowing what agents are good for, drawn from write-in responses to the survey.
High developer trust in AI agents doesn’t remove the need for manual review
91.1 percent of respondents describe agents as having improved or dramatically increased their productivity. Just 4 percent report no impact, and 1.8 percent describe a decline.
Trust follows a similar shape. 85.5 percent of respondents trust agent outputs at least somewhat, made up of 24.7 percent who completely trust them and 60.8 percent who somewhat trust them; only 4.0 percent distrust them outright.
Successful teams show more confidence than everyone else. They are 6.1 times more likely to completely trust outputs, at 28.4 percent against 4.7 percent, and 1.6 times more likely to somewhat trust them, at 90.8 percent against 57 percent. Temporal users land at a similar level of confidence, 1.1 times more likely than the wider sample to trust their agents at least somewhat, 95.7 percent against 84.7 percent.
That confidence stops short of full autonomy. 39.5 percent of respondents point to security concerns as the top block on fully autonomous agents, above every other factor. The report frames the block as an old outsourcing problem in new form: engineers need certainty that delegated work was done correctly, and short of time-consuming testing and mathematical proofs, manual checking remains the only way to get it.
With that in mind, the report draws a line between trusting an agent enough to proceed and trusting it enough to put a name on the commit, despite 85.5 percent of respondents trusting outputs to some degree.
84.5 percent of respondents believe their team uses agents better than competitors do, a distribution the report calls statistically unlikely given that respondents cannot all occupy the top of the field. Leadership drives part of that confidence. CEOs, managers, vice presidents, and directors are 1.4 times more likely than engineers to call their organisation “very successful” at using agents, 43.7 percent against 32.4 percent, and the report puts the leadership-to-engineering gap at 35.1 percent overall.
Production speed leaves QA cycles catching up
51.3 percent go from an AI-generated prototype to production-ready code in hours or faster, and 77.1 percent reach that point within days.
The speed doesn’t track only to startups: 33.4 percent of the sample works at companies with 1,001 or more employees, and 14.6 percent at companies with 5,001 or more. Responses within that largest cohort split into two camps. Most deploy at the pace of smaller companies. A slower fifth (18.5%) takes months, against 5.5 percent of everyone else.
Successful teams gain only a 1.2 times speed advantage over everyone else, 78.9 percent reaching daily-or-faster speed against 67.4 percent. The report notes that several engineers named time to think, not tooling speed, as their main constraint. That points to acceleration reaching every team at close to the same rate; outcomes depend on what a team does with the time it saves.
Tracking state and debugging top the list of blockers
Respondents named tracking state as the top blocker to using agents more, at 35.7 percent, ahead of debugging in second place and managing token or compute costs in third.
Successful teams give more answers per respondent, pointing to a wider set of issues or sharper awareness of them. They are somewhat more likely to mention tracking, retries, unrelated business logic, and infrastructure, and somewhat less likely to mention orchestrating, isolating or managing costs.
41.1 percent of respondents encounter agent issues daily or more, and 16.4 percent hit them hourly or more. Issue frequency does not track to the number of agents a team runs. Teams operating dozens of agents encounter no more trouble than teams running five, though the report flags that definitions of an issue vary by team and closer attention tends to surface more of them.
Successful teams are just as likely to encounter issues “daily or more.” They are 1.7 times more likely to call those issues continuous, a higher frequency than daily; everyone else is 1.3 times more likely to stop at the daily mark alone.
YouTube ranks as the top destination for troubleshooting help when something breaks, ahead of GitHub and Stack Overflow, platforms that let users copy code directly. The report reads this as a preference for watching someone work through a result on screen over copying a pattern from text.
Successful teams search more widely when trouble hits. They check 4.1 sources on average against 3.5 for everyone else, and they are twice as likely to turn to an AI tool for help, 40.4 percent against 19.8 percent, and 1.9 times more likely to turn to the social media platform X, 31.6 percent against 16.3 percent.
Cost and tooling choices among agent-heavy teams
79.8 percent of respondents say token and compute costs limit their progress with agents. Successful teams are 1.3 times more likely to call cost at least somewhat of a factor, 83.1 percent against 65.7 percent. Everyone else is more likely to remain unsure, 22.1 percent against 7.3 percent.
ChatGPT or its API ranks as the most-used AI tool, followed by Microsoft Copilot, Google Gemini, and Claude. OpenAI’s first-mover position and wider product range keep it ahead.
OpenAI-linked products account for most of the list once its own tools are counted alongside services built on its models, such as Copilot. The report attributes Copilot’s and Gemini’s reach to their default placement inside existing work tools, with Claude sitting just above ChatGPT Codex in fourth place.
The average team uses 3.5 of these tools, and successful teams use 1.3 times more, 3.6 against 2.7. Successful teams also favour specific platforms more heavily. They are more likely to use AWS Bedrock, 23.1 percent against 8.1 percent, and Azure OpenAI Service, 30.3 percent against 18.6 percent.
GitHub Copilot shows a similar gap, 32.7 percent against 15.1 percent, and Google Gemini or Vertex AI reaches 44.2 percent against 29.1 percent. LangChain or LangGraph usage more than triples among successful teams, from 2.3 percent to 7.3 percent.
OpenAI products remain the most commonly-selected agent framework. Microsoft’s framework gained ground last year, alongside AWS Strands and CrewAI. Successful teams are 1.7 times more likely to build agents on the OpenAI Agents SDK, 51.1 percent against 30.2 percent, and 1.9 times more likely to use Temporal itself, 9 percent against 4.7 percent.
Interest in orchestration tools is changing only in pockets. Interest in AWS Step Functions, LangGraph, and Airflow is falling. Interest in Camunda, Inngest, and Prefect is rising in its place. Temporal, Dagster, and Orkes held roughly constant. Successful teams show less appetite for changing their orchestration approach than everyone else, consistent with leaving a working setup alone.
Custom-built software adds new territory for testing teams
92.3 percent of respondents have tried building an application in-house that their team would previously have bought as a product, and 25.6 percent count the attempt as a success with a big impact.
Successful teams are 8.5 times more likely to log a rebuild with a big impact, 29.7 percent against 3.5 percent. Everyone else is 2.9 times more likely to have skipped the attempt or seen it fail, 17.0 percent against 6 percent. Temporal users are 1.6 times more likely to count rebuilding instead of buying as a success with a big impact, 39.1 percent against 24.4 percent.
The report raises a question for future study without answering it: how development and operations teams will manage a growing set of custom, agent-generated tools that don’t automatically work together, and whether that changes onboarding time for new hires. Both remain open items alongside the survey’s data on testing, debugging, and trust.
See also: AWS adds OpenAI’s GPT-5.6 to Kiro’s agentic coding workflow
Want to learn more about cybersecurity from industry leaders? Check out Cyber Security & Cloud Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and is co-located with other leading technology events including the AI & Big Data Expo. Click here for more information.
Developer is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.