TL;DR: according to the State of Development 2026 report published by Temporal on August 26, 2026, 80.8% of the engineers surveyed use AI agents at least once a day, up from 47.3% a year earlier; and 92.3% of teams have tried to rebuild in-house software they previously bought as SaaS.
This number is not just about developers. It concerns every company that buys software, because it describes the new economics of the trade that produces it. When code production speeds up this much, two things move at once: the price of custom software drops, and the gap widens between teams that ship reliable systems and teams that ship demos.
Key takeaways
- Temporal's State of Development 2026 report, published on August 26, 2026, found that 80.8% of the 554 engineers surveyed use AI agents at least once a day, up from 47.3% a year earlier.
- 92.3% of the teams surveyed have tried to build in-house applications they previously bought as SaaS; only 25.6% succeeded with significant impact.
- 49.1% report agents in production or essential to their deployment, yet 41.1% run into agent-related issues at least once a day.
- A May 2026 Stack Overflow survey adds nuance: 63% of professionals rarely or never let an agent run without human oversight.
What Temporal's August 26 report actually measures
Temporal Technologies, the company behind an orchestration platform used to make distributed applications reliable, published the second edition of its State of Development report on August 26, 2026, dedicated this year to AI agents. The survey was run by Qualtrics from April 29 to May 25, 2026 among 554 qualified respondents: engineers, architects, data leads and IT executives, two-thirds US-based and one-third from the UK and EMEA, in companies ranging from under 50 to over 5,000 employees.
The usage numbers are stark. 80.8% of respondents use AI agents at least daily, up from 47.3% a year earlier, a relative jump of 70.8%. Teams run a median of 5 agents, with an average of 10.7 that betrays the existence of heavy users. And 51.3% go from prototype to production code in hours or less, according to the announcement relayed by MarTech Series on August 26, 2026.
One methodological caveat, which the report does not hide: the sample is made up of professionals who already use AI agents. The 80.8% figure therefore describes usage intensity among converted teams, not adoption across the whole market. For the broader picture, the Stack Overflow survey published on May 27, 2026 among 1,100 developers reports lower figures: 59% use agentic AI, including 37% daily. Same trend, same slope.
Why are 92.3% of teams rebuilding what they used to buy?
Because the cost of building has collapsed. This is the most consequential number in the report: 92.3% of the teams surveyed have tried to build applications themselves that they previously bought as SaaS subscriptions. The classic build-versus-buy trade-off, which almost always tilted toward buying for a mid-sized company, is rebalancing, because a small internal tool that used to take three months of development now takes a few weeks.
The opposite number matters just as much: only 25.6% of those attempts delivered significant impact. Three attempts out of four produce no notable effect. Building has become easy; succeeding has not. The report shows where the difficulty lives: 41.1% of respondents hit agent-related issues at least daily, 79.8% cite compute cost control as a challenge, and state tracking tops the list of technical obstacles. Samar Abbas, CEO of Temporal, sums it up: the teams pulling ahead are the ones that trust their systems, because they have solved for state, cost and reliability.
The table below sums up the numbers worth keeping.
| Indicator | Figure (2026) | Reading |
|---|---|---|
| Daily AI agent usage | 80.8% (47.3% in 2025) | The agent has become the default work tool |
| Agents in production or essential | 49.1% | Half of teams are past the pilot stage |
| Teams that tried replacing a SaaS | 92.3% | The build-versus-buy question is open again |
| Attempts with significant impact | 25.6% | Building is easy, succeeding is still rare |
| Agent issues at least daily | 41.1% | Reliability is the real work |
What does this change for a small or mid-sized business?
First, the price and lead time of custom software are falling. A line-of-business tool, a CRM shaped around your processes or an automation that used to require an agency budget are now within reach of a company with 10 to 100 employees, because the provider building them produces faster. If a 2026 quote looks like a 2023 quote, in price and in timeline, it deserves a question.
Second, the criteria for choosing a provider are shifting. When anyone can produce a convincing demo in two days, the demo no longer proves anything. The report's numbers say it plainly: the gap between the 25.6% who succeed and the rest is not decided by code production speed, but by everything around it. Who monitors the tool while it runs? What happens when a step fails at 3am? What does the tool cost to operate, not just to build? Those are the questions to ask before signing, and a serious provider should answer them without improvising.
Third, the lesson of the three-in-four failure rate applies internally too. The fact that 92.3% of engineering teams are trying to replace their SaaS does not mean your company should. A $50-a-month subscription that does the job remains unbeatable; custom software earns its keep when the off-the-shelf tool forces you to bend your processes, when subscriptions pile up, or when your data deserves to stay in-house. Sometimes the right answer is to keep your current tool, and honest advice should be able to tell you so.
Finally, there is supervision. The May 2026 Stack Overflow survey shows that 63% of professionals rarely or never let an agent run fully autonomously, and 60% block agents from making unapproved system changes. That is not timidity, it is engineering: a useful agent is one that is monitored, logged, and confined to a clear perimeter. I described this gap between prototype and production in the article on the death valley of AI agents; the Temporal report puts numbers on it.
How this connects to my day-to-day work
These numbers describe my practice. AI is my production accelerator: it is what lets me ship tools to an SME in a few weeks, from €400 for a sprint, that would have required an agency budget three years ago. But what makes them last is the work of the 25.6%: systems that log what they do, alert when a step fails, and repair themselves when possible. The IA Brew newsletter, 93 n8n nodes running on their own, works because every critical node has an error path; not because the code was written fast.
The Temporal report does not say AI agents are replacing developers. It says the trade has changed tools, and that value has moved from writing code to keeping what runs reliable. For a business owner, the useful question is no longer "how much does development cost?" but "who answers when it breaks?".
Frequently asked questions
What does the Temporal 2026 report say about AI agent adoption?
Temporal's State of Development 2026 report, published on August 26, 2026, found that 80.8% of the 554 engineers surveyed use AI agents at least once a day, up from 47.3% a year earlier. The survey was run by Qualtrics from April 29 to May 25, 2026 among professionals already using AI agents, across the US, UK and EMEA.
Are teams really replacing their SaaS tools with custom software?
According to the Temporal 2026 report, 92.3% of the teams surveyed have tried to build in-house applications they previously bought as SaaS, but only 25.6% succeeded with significant impact. Building has become accessible; success depends on reliability, state management and cost control, not on the initial demo.
Are AI agents reliable in production?
Not yet without supervision. The Temporal 2026 report found that 41.1% of respondents run into agent-related issues at least daily, and a May 2026 Stack Overflow survey showed that 63% of professionals rarely or never let an agent run fully on autopilot. Human oversight and reliability engineering remain the norm.