By ByteBot
Most engineers have crossed the threshold. AI agents are now a daily work tool, not a weekend experiment. Temporal’s freshly released 2026 State of Development report — surveying 554 engineers and engineering leaders across the US and UK — puts the number at 80.8%, up from 47.3% a year ago. That is a 70.8% jump in frequent agent use in twelve months. Read past the headline, though, and a more complicated picture emerges: the same engineers who use agents every day cannot agree on how to run them reliably, and the gap between teams that have figured it out and teams that haven’t is accelerating.
The State Tracking Problem Is Bigger Than Model Quality
The intuitive assumption is that hallucinations — wrong answers confidently delivered — are the primary obstacle to broader agent adoption. The Temporal survey says otherwise. When asked what blocks them from using more AI agents, 35.7% of respondents named state tracking as the top concern. Debugging came second. Cost came third. Model quality barely registered as the primary complaint.
This matters because it reframes where the work needs to happen. If your agents keep failing, the culprit is probably not the model — it’s the infrastructure around the model. When an agent fails mid-run, state is lost and the entire job restarts from scratch. For long-running workflows — overnight research tasks, multi-step code generation pipelines, API integrations spanning dozens of tool calls — that is not a minor inconvenience. It is a production blocker that no amount of prompt engineering fixes.
32% of Organizations Are Already Skipping Software Purchases
The economic signal is harder to ignore than the adoption survey. McKinsey’s State of AI 2026 found that 32% of organizations decided against buying at least one software product or feature because they could build it in-house with agentic coding tools instead. Tech leads the shift at 41%, followed by healthcare (39%) and professional services (38%). This is the build-vs-buy question getting answered at scale, and the answer is increasingly “build” — not because building is easier, but because agents have made it fast enough to be economical.
McKinsey’s verdict: enterprise AI is finally on the road to ROI. That is a sharp turn from the skepticism that dominated 2025. The organizations driving that shift are the ones that got the infrastructure right first.
The Reliability Math Is Unforgiving
Here is the problem that doesn’t make it into the demos. If each agent in a three-agent chain succeeds 70% of the time individually, the chain succeeds only 34% of the time overall. Compound that across longer workflows and the numbers collapse fast. LangChain’s State of Agent Engineering survey of 1,300 practitioners found that 57% have agents in production — but 48% skip offline evaluations entirely and 63% skip online monitoring. You cannot debug what you cannot see, and most teams cannot see much.
The numbers from Deloitte are starker: 89% of AI agent pilots never reach production. Inngest’s AI in Production 2026 report pinpoints the leading complaints: agent framework abstractions make failures harder to trace (26%), and poor support for long-running or stateful workflows (19%). The theme is consistent across every survey. The model is not the weak link. The plumbing is.
The Gap Is Widening
Temporal’s report flags a sharp acceleration in adoption alongside a widening performance gap between teams that have resolved the reliability challenges and those still catching up. That gap will not close on its own. Teams that built observability into their first agent prototype, chose durable execution infrastructure before they needed it, and established evaluation pipelines before shipping — those teams are shipping agents that work. Teams that skipped those steps are still rebooting failed workflows by hand.
If 80.8% of engineers are using agents daily but only 49.1% say agents are core to how they ship, that delta is the real story. Thirty-one percent of engineers trust agents enough to use them every day but not enough to make them load-bearing. That is the problem the industry needs to solve — and the data makes clear it’s an infrastructure problem, not a model problem.
