AI Blog

85.5% trust the agent. 41.1% debug it every day.

Temporal surveyed 554 engineers in April and May 2026 and found daily agent use at 80.8%, up from 47.3% a year earlier, with 91.1% reporting improved productivity and 85.5% trusting agent output at least somewhat — alongside 41.1% hitting agent-related issues daily or more and 9.0% continuously. Both sets of numbers are probably accurate, and together they describe a failure rate nobody would accept from a database. The report reads the gap as a state-tracking problem, which is a durable-execution vendor’s reading of a durable-execution question. The more useful reading is that the error handler is a person, and no dashboard has a line for them.

By Agentic AI Wiki 14 min read

Put two findings from the same survey next to each other and the week's most-quoted adoption statistic stops being good news: 85.5% of engineers say they trust agent output at least somewhat, and 41.1% say they hit agent-related issues daily or more, with 9.0% saying continuously. There is no reading of those two numbers under which the trust is well calibrated. What makes the combination stable rather than a crisis is the thing the survey cannot see — the error handler is a person, absorbing failures one at a time, and no dashboard anywhere has a line for the hours they spend doing it.

What the report actually says

Temporal published its 2026 State of Development Report on AI agents in late August. The survey ran from 29 April to 25 May 2026 across the US and UK and, after low-quality responses were removed, covers 554 software engineers, architects, infrastructure contributors and engineering leaders. The headline is adoption, and it is genuinely steep.

FindingFigureWhat it is measuring
Use agents daily or more80.8%, up from 47.3%Self-reported frequency — a 70.8% relative jump in a year
Agents "improved" or "revolutionised" productivity91.1%Perceived value, unbounded and uncompared
Trust agent output at least somewhat85.5%A disposition, not a delegation decision
Hit agent-related issues daily or more41.1% (9.0% continuously)Encountered failures, undefined and unpriced
Top blocker to using agents moreTracking state, 35.7%What engineers name, then debugging, then cost
Sentiment and failure findings from the same survey Horizontal bar chart on a nought to one hundred per cent scale, split into two groups. The sentiment group shows agents improved or revolutionised productivity at 91.1 per cent, trust agent output at least somewhat at 85.5 per cent, and daily agent use at 80.8 per cent with a marker at the 47.3 per cent figure from a year earlier. The failure group shows agent-related issues daily or more at 41.1 per cent, tracking state named as the top blocker at 35.7 per cent, and agent-related issues continuously at 9.0 per cent. All six figures come from the same 554 respondents. One survey, 554 respondents (per cent) 0 25 50 75 100 HOW ENGINEERS FEEL ABOUT AGENTS Productivity improved 91.1 Trust output at least somewhat 85.5 Use agents daily or more 80.8 47.3 a year earlier WHAT AGENTS DID TO THEM LAST WEEK Agent issues daily or more 41.1 Top blocker: tracking state 35.7 Agent issues continuously 9.0 Surveyed 29 April – 25 May 2026, US and UK, self-reported.
The top three are how engineers feel about agents. The next two are what agents did to them last week.

Take the report's framing at face value first, because it is defensible. Its argument is that adoption outran infrastructure: engineers picked up agents faster than their organisations built the systems to run them, and the teams pulling ahead are the ones that solved state, cost and reliability. That is a coherent story and the blocker data supports it. It is also, precisely, the story a durable-execution company is in business to tell, and the survey was designed by people who already believed it. That does not make it wrong. It makes it worth checking against the numbers that were not the headline.

"Trust" is the wrong variable, and the survey inherits the mistake

Trust as a scalar against trust as a per-action delegation Three columns. The first, trust as a scalar, is what a survey asks: do you trust the output, answered on one agree-disagree scale that aggregates every action an agent could take and cannot separate reading a file from issuing a refund. The second, trust as a per-action delegation, is the decision a team actually makes: a list of which actions run unreviewed, which require approval and which are never delegated, each carrying its own reversal cost. The third column notes that the first is a disposition that costs nothing to hold while the second is a policy with a blast radius, and that a team can score high on the first while having never written the second down. Trust as a scalar "do you trust the output?" one agree–disagree scale reading a file = issuing a refund WHAT THE SURVEY ASKED costs nothing to hold, and predicts nothing Trust per action runs unreviewed needs approval first never delegated WHAT THE TEAM DECIDES each line carries its own reversal cost The gap 85.5% answered the first most teams never wrote the second 41.1% live in the difference MEASURE INSTEAD review rate per action class, and change rate when reviewed
The first column is cheap to answer. The second is the one with a blast radius.

"Do you trust agent output?" aggregates over every action an agent might take, and the aggregate has almost no operational content. Trusting a summary of a stack trace and trusting a database migration are different decisions with different reversal costs, and an engineer who answers "somewhat" is describing a feeling about a category. The decision that governs blast radius is not a scalar at all: it is a list of which specific actions may run unreviewed, which need approval, and which are never delegated — the ladder that autonomy levels lays out, and the thing most teams have never written down.

This matters because the two numbers only look contradictory under the scalar reading. Under the per-action reading they are perfectly consistent and considerably more worrying: engineers have high trust in the class of output they review, are shipping faster than ever, and are hitting failures daily in the actions they no longer review. The survey cannot tell those apart, and neither can a team that has only ever discussed trust as a mood.

The practical replacement for the question is two lines in a run record. For every action class an agent can perform, what fraction of runs was reviewed by a human before the action landed, and what fraction of reviewed actions was changed? A high review rate with a near-zero change rate is a control you can safely loosen. A low review rate with a high change rate is where your 41.1% lives.

Engineers named the blocker they can see

State tracking at 35.7%, then debugging, then cost. Read that as a ranking of visibility rather than of cause. State is what you are staring at in the debugger when a run fails halfway and cannot be resumed; it is legible, nameable and has a product category attached to it. That is exactly the kind of answer a survey elicits, and it is why "tracking state" beats answers like "we never defined what done means for this task" — which is upstream of it, harder to name in a checkbox, and considerably more likely to be the actual defect.

State really is a hard problem and the durable-execution engines really do solve part of it; the wiki's own comparison of where an agent's transcript lives covers what they do and do not fix. But notice what a resumable run buys you and what it does not. It buys you the ability to continue a run that crashed. It does not tell you whether the run should have continued, whether the work it had already done was correct, or whether resuming it will repeat a side effect. Those are eval and idempotency questions, and they are the ones a daily failure rate is usually made of — see idempotency and retries, where four stacked retry sources mean every write tool fires twice unless you construct exactly-once deliberately.

The tell that the blocker ranking is about visibility: cost placed third, in a population where 41.1% report daily failures. A daily failure rate is a cost — the retried run, the discarded output, the engineer's afternoon — and it ranks below cost because none of it arrives as a line item. The provider bill is legible. The labour absorbing the failures is not.

The error handler is a person

Where an agent failure terminates, compared with a service failure Two horizontal paths. The upper path, a conventional service, runs from a failure through a code error handler, then a retry or fallback, then an alert, and ends in an incident record and error budget that a team reviews. The lower path, an agent workflow, runs from a failure to the engineer at the keyboard, who rewrites the prompt or reruns the task, and terminates there: no alert, no incident record, no error budget entry. A panel underneath explains that this is why a 41 per cent daily failure rate can persist without appearing on any dashboard, and proposes a single boolean field on the run record recording that a human intervened. A CONVENTIONAL SERVICE failure timeout, 500, bad row error handler retry, fallback alert a pager, a threshold incident record, error budget reviewed, trended, argued about AN AGENT WORKFLOW failure wrong patch, lost state the engineer notices it directly rerun, rewrite, redo about a minute nothing is emitted no alert, no record, no budget line A handled error is still an error. This one is handled by the most expensive component in the loop, which is why 41.1% can be true and invisible at the same time. The cheap fix: one boolean on the run record — a human intervened — plus the task class. A week of that ranks task classes by intervention rate, and the daily failure rate stops being evenly spread.
Both paths end. Only one of them ends somewhere a team can see.

Here is the mechanism that makes 41.1% survivable. In a conventional service a failure hits a code path — a retry, a fallback, an alert — and terminates in an incident record and an error budget somebody reviews at the end of the sprint. In an agent workflow the failure is usually caught by the engineer sitting in front of it, who rewrites the prompt, reruns the task, or does the work by hand, and the path ends there. Nothing is emitted. No alert fires because the human already handled it, which is the definition of a handled error and also the reason it never becomes data.

This is why the aggregate can look so strange. Each individual failure is cheap — a minute, a rerun, a slightly different prompt — and cheap-and-frequent is the profile that never gets escalated. It is also the profile that quietly sets your team's throughput ceiling, because the minutes are real and the person spending them is the most expensive component in the loop. The economics page on the cost of human review makes the same argument from the other end: the reviewer is ten to fifty times the token cost and is the one line that does not shrink when the model gets better.

The fix is not to eliminate the failures, which is not on offer. It is to make the handling visible, and that costs almost nothing: one field on the run record for "a human intervened", set by whatever surface the intervention happened on. Once a week's runs carry it, you can rank task classes by intervention rate and discover that the daily failure rate is not evenly spread — it is two or three task classes generating most of it, which is an actionable finding and the one thing a survey can never give you about your own system.

Throughput went up. Review did not.

The report's most under-discussed number is that 51.3% now go from prototype to production-ready code in hours or faster, and 26.9% in minutes or faster. Set that beside the daily failure rate and the shape of the risk becomes clear: the generation step got dramatically faster and nothing downstream of it did. Review capacity is unchanged, because it is bounded by human attention, and the number of changes arriving at it went up several-fold.

That is the same bottleneck the background coding agents playbook argues about pull requests — an agent that opens twelve a day adds nothing if the team merges four — restated at the level of a whole organisation. A queue whose arrival rate multiplies and whose service rate is flat does not degrade gracefully; it either grows without bound or the service rate silently drops, which in a review queue means approvals get shallower. Nobody decides to review less carefully. It happens because the alternative is a queue that never empties.

The hiring findings sit oddly alongside this and are worth one paragraph of caution: 56.7% believe it will be harder for junior engineers to find jobs and 45.5% say the same of senior engineers, while only 26.4% of companies report slowing or stopping hiring. Engineers' expectations have moved considerably further than their employers' behaviour has. Both numbers are self-reported by the same population, so the gap is a fact about sentiment, not about the labour market, and it should be read that way.

How much of this to believe

The caveats are ordinary and they matter. This is a vendor survey, which shapes what was asked as much as how it is framed. The population is self-selected, 554 respondents across two countries, and every figure is self-reported — including "productivity improved", which is measured against no baseline and by people who chose to answer a survey about agents. "Agent-related issue" is undefined, so it spans a hallucinated import and a production outage. And "daily agent use" in a survey of software engineers is heavily a coding assistant, which is a different animal from an autonomous agent taking actions in production; do not read the 80.8% as deployment.

What survives all of that is the internal comparison, because both figures come from the same instrument and the same respondents. Whatever bias inflated the trust number would inflate the failure number too, and the failure number is high anyway. A population that reports failing daily and trusting anyway is telling you something real about how agent work is currently absorbed, and it is the finding worth taking away from a report whose headline is adoption.

The one number to steal for your own team is not in the report at all. Take a week of runs, count what fraction ended with a person intervening, and split it by task class. If the answer resembles 41%, you now know which two or three classes to fix — and unlike a trust score, it moves when you fix them.

FAQ

Is a 41.1% daily failure rate actually bad, or is that normal for new tooling?

It is unremarkable for tooling you drive by hand and unacceptable for anything acting unattended, and the survey cannot tell you which of those it measured. That is why the interesting version of the question is per action class rather than in aggregate. A code-completion suggestion failing daily is a rounding error on your day; a scheduled agent writing to a system of record failing daily is an incident you have not opened yet.

Does the report show that durable execution is what teams need?

It shows that state tracking is what engineers name first when asked what blocks them, which is a real signal and a partial answer. Resumability fixes runs that crash; it does not fix runs that complete incorrectly, and it does not stop a resumed run from repeating a side effect unless you also built idempotency. If your daily failures are wrong outputs rather than lost progress, a durable-execution engine will not move the number.

Why does trusting agent output "at least somewhat" tell us so little?

Because it aggregates across actions with wildly different reversal costs and asks about a disposition rather than a delegation. The operational question is which actions may run without review, and it has a different answer for reading a file, drafting a message, and issuing a refund. A team can be at 85% on the survey question and have never written the second list down.

Should I be worried that adoption jumped 70.8% in a year?

Not by itself — a fast-rising adoption curve is what a useful tool looks like. Worry about the pairing: adoption rose, self-reported time-to-production collapsed to hours or minutes for half the respondents, and review capacity did not change, because it is bounded by people. That combination is where approvals get shallower without anyone deciding to make them shallower.

What is the cheapest instrumentation that would answer this for my own team?

One boolean per run for "a human intervened", written by whatever surface the intervention happened on, plus the task class. That is enough to rank task classes by intervention rate within a week and to tell a genuinely reliable agent apart from one whose failures are being quietly absorbed. Everything else in agent observability is more useful, and none of it is cheaper.

Further reading

On this wiki:

Sources: