80.8 percent of engineers use agents daily, 41.1 percent hit problems every day
Per the AI agents report Temporal published on August 26, 2026, 80.8 percent of respondents use agents daily or more often. In the same group 41.1 percent say they run into agent related problems daily. The survey covered more than 550 engineers and engineering leaders.
A large majority of engineers use agents daily
Temporal published the second edition of its state of development report, this one devoted to AI agents, on August 26, 2026.
Per the report, 80.8 percent of respondents use AI agents daily or more often. A year earlier that figure was 47.3 percent, a 70.8 percent relative increase.
The survey ran from April 29 to May 25, 2026 and carries responses from more than 550 software engineers, architects, infrastructure contributors and engineering leaders. Two thirds of respondents are US based, one third from the UK and EMEA.
Adoption is fast, the infrastructure lags
The report's core finding is the gap between adoption speed and operational maturity. Some 91.1 percent of respondents say agents have improved or revolutionized productivity.
In the same group, 41.1 percent say they hit agent related issues daily or more often. Another 9 percent describe the problems as continuous.
On trust the picture is mixed: 85.5 percent trust agent output at least somewhat. Running into problems and trusting the output can coexist.
On scale the median respondent runs 5 agents while the average reaches 10.7. That gap points to a small number of organizations operating a large number of agents.
On speed, 51.3 percent say they move from prototype to production code within hours, and 26.9 percent do it in minutes or faster.
Temporal CEO Samar Abbas says in the report that engineers adopted AI agents faster than most teams built the infrastructure to run them reliably.
The goal is not avoiding failure, it is managing it
The assessment in this section is ours. The most striking pair in the report is 91.1 percent against 41.1 percent. The same group says productivity rose and that it hits problems every day.
That is not a contradiction, it is the normal state of an immature technology. Agents are valuable not because they do not fail, but because their failures can be absorbed.
The real question is whether that absorption is built. If a wrong step by an agent can be undone, daily problems are an acceptable cost. If it cannot, the same rate is a serious risk.
The gap between a median of 5 agents and an average of 10.7 says the same thing: as the count rises, agents become an architecture problem. You can watch five by eye, you cannot watch fifteen.
The share moving from prototype to production within hours is a separate warning. Shipping fast is good, but if the question of who approved what went to production stays unanswered, speed stops being an advantage.
What it means for businesses in Türkiye
The assessment below does not appear in the source, it is our reading. Respondents are weighted to the US, UK and EMEA, and no Türkiye specific breakdown is published.
The findings still describe the local picture well. The most common pattern we see in the field is this: a team tries an agent, it works, it gets wired into a process immediately, and when that process should be stopped is never written down.
For small and mid sized businesses the practical first step is narrowing scope. Let the agent do one job end to end rather than ten jobs halfway. A narrow scope makes problems visible and makes undoing them easier.
The second step is an approval gate. Putting a human approval in front of every step that produces a consequence looks like lost speed, but given the daily failure rate it is cheap insurance.
The third step is logging. A system that does not record what the agent did in readable form cannot be diagnosed when something breaks. This is the item most often skipped in Türkiye.
The fourth is expectation management. A figure like 91 percent productivity improvement is a survey result; what it is worth in your own business cannot be known without measuring your own data.
The UNALSOFT take
In agentic AI projects the first thing we write down is not what the agent can do but what it cannot. Keeping scope narrow, granting permissions by name and putting approval in front of consequential steps are not three separate choices for us, they are one design decision. The 41.1 percent in this report is exactly why: using agents does not mean avoiding problems, it means being able to close them cheaply. Another thing we see in the field: teams stand an agent up quickly but never build a record they can read afterwards. The time lost when something breaks far exceeds the time building that record would have taken. Speed is a real advantage only together with reversibility.
Sources
Temporal, The State of Development 2026 · MarTech Series, the 70.8 percent leap in the report
Can your agent's mistakes be undone?
Let us write the scope, approval gate and logging plan before the system is built.