Adoption Is No Longer the Interesting Question
80.8% use agents daily or more, against 47.3% a year ago. The jump is large enough to warrant checking the definition: the report defines an AI agent as an LLM instance capable of taking multi-step actions on its own, and asks about frequency of use rather than depth of dependence. "Uses one daily" and "could not work without one" are different claims. The more interesting finding is the great separation: the gap between teams running agents well and teams still catching up is widening, and the dividing line is not how many errors they hit. Temporal found that successful teams face the same rate of errors while being less worried and less stressed, and that their advantage compounds — they use more tools, search more widely for troubleshooting help, and build more of their own developer applications. Put differently, what separates teams is not whether agents fail. They all do. It is whether there is machinery to catch the failure.
Three Words Where Things Stall: State, Cost, Reliability
Abbas's framing is worth quoting directly: engineers have adopted AI agents faster than most teams have built infrastructure to run them reliably, and the teams pulling ahead are those that have solved for state, cost and reliability. Each word maps to a concrete problem. State is what happens when an agent dies halfway through — resuming versus starting over. Cost is whether token spend across different configurations of the same task can be compared side by side. Reliability is whether a multi-step workflow behaves predictably when one step fails. Temporal sells durable execution, so this framing has an interest behind it, but these three are genuinely the first walls you hit moving agents from demo to production.
Conditions to Read the Data With
The sample is 554 (650 raw responses, reduced after removing low-quality ones), two-thirds US and one-third UK/EMEA, spanning companies from under 50 to over 5,000 employees, with the largest concentration at mid-sized firms (29.2%). Roles: engineer/AI engineer 25.6%, VP or director of IT 13.9%, data engineer 11.9%. Industries: software 37.5%, manufacturing and professional services 9.7% each. The survey ran through Qualtrics, screening for people currently using AI agents — which matters, because it measures frequency change within users, not penetration across all engineers. The other condition is timing: data was collected April 29 to May 25, 2026, more than three months before publication. In this field three months is enough for a toolchain generation to turn over, so read it as a directional trend rather than a current snapshot. The practical implication: if your team's agents are stuck at "great individually, falls over in production," this report says stop looking at the model for the answer. Look at how state is persisted, how failures retry, and how spend is attributed.