80.8% run agents daily. State tracking is the bottleneck

Adoption stopped being the story this year. In Temporal's 2026 State of Development report, 80.8% of surveyed engineers now use AI agents daily or more, up from 47.3% a year earlier. The number that should change how you build is further down: the single biggest blocker to using agents more is tracking state, not model quality.

What actually happened

Per Temporal's report, the survey ran April 29 to May 25, 2026, and reports 554 respondents after filtering, split roughly 67% US and 33% UK/EMEA. Daily-or-more agent use came in at 80.8%, against 47.3% a year prior — a 70.8% relative jump.

Deployment is further along than the "still experimenting" narrative suggests: 49.1% say agents are in production or core to how they ship, with 21.8% in that "core" bucket.

Then the friction. 35.7% named tracking state as the top blocker to expanding agent use, ahead of debugging and cost management. Trust is broad but shallow — 85.5% trust agent output at least somewhat, while only 24.7% trust it completely, and 41.1% hit issues daily or more often.

One disclosure, because it matters: Temporal sells durable execution. A survey finding that durable state is the hard part is a survey run by the company that sells durable state. Read the numbers, discount the framing. The finding still matches what we see in client stacks.

Why agent reliability matters for your business

An agent that works in a demo and fails in production usually fails the same way: something died mid-run. The API timed out on step four of seven, the process restarted, and now nobody — including the agent — knows whether step three already charged the card or sent the email.

That is not a model problem and no upgrade fixes it. It is the same distributed-systems work that has always sat under payments and order processing, and it is unglamorous: make every step idempotent so a retry is safe, persist state outside the agent process so a crash doesn't erase the run, and keep a durable log of what was attempted and what actually happened.

The 41.1% hitting issues daily is the tell. Those teams are not short a smarter model. They are short a resume point and an audit trail — which is also what turns "the agent did something weird" from a mystery into a ticket.

Practical order of operations: before you add the fifth agent, make the first one survive a restart. Pick your highest-volume workflow, kill it halfway through on purpose, and see what it does. Whatever breaks is your actual roadmap. We've argued the same thing about putting an audit trail under agents — this survey is 554 engineers saying it out loud.

Sources: Temporal, Developer Tech.