The Code Step Collapsed. Your Delivery Probably Didn't.

Half the engineers in Temporal's State of Development 2026 go from AI prototype to production-ready code in hours or faster. A quarter say minutes. If that were the whole story, delivery lead times across the industry would have fallen off a cliff this year.

‍

Same survey, further in: 41.1% run into problems with their agents daily or more often, 16.4% hourly or more. The single most common blocker to running more agents is tracking state, at 35.7%. Four in five say token and compute cost limits how far they can push.

‍

The build step got fast. The system around it didn't.

Read the sample before you quote the headline

Temporal surveyed 554 engineers and engineering leaders between April 29 and May 25, 2026, all of them already using AI agents, about two-thirds in the US. That's a survey of adopters, not of the profession. So 80.8% daily use means four in five people who already use agents use them constantly, which is a different claim than four in five engineers. Same for the 91.1% who say agents improved or revolutionized their productivity.

‍

The internal comparisons are where the value sits. When the report contrasts teams that call themselves successful against everyone else, both groups hit agent problems at the same rate. Success wasn't fewer failures. It was better handling of them.

The number that made us stop

84.5% believe they're better at using AI agents than their competitors. Most of them are wrong, by arithmetic.

‍

That's an easy joke, and it hides the real finding. Ask an engineering leader how their agent program is going and they answer from the fastest squad in the building, because that's the demo they've seen. Nobody's comparing against a competitor's lead time, because nobody has that number. They're comparing against a feeling.

‍

Meanwhile the same population reports daily breakage, a state management problem they can't get past, and costs they can't forecast. Confidence is running well ahead of evidence, and evidence is the only thing that tells you which investment to make next.

Where the constraint went

AI didn't remove your bottleneck. It relocated it.

‍

The way we model delivery is one loop with eight stages and two gates: clarify, spec, approve, build, validate, accept, release, measure. Agents now help with most of those stages. Two stay stubbornly human. Somebody approves the spec, and somebody accepts the build.

‍

Your cadence equals your slowest approval loop. If your product organization makes scope decisions once a week, you ship on a weekly rhythm no matter how many minutes the code generation takes. The 51.3% figure describes the middle of that loop. The two ends set the pace, and the report doesn't measure them, because most teams don't either.

‍

Pull your last twenty stories and split each one into days worked versus days waiting on a decision. If waiting dominates, buying more seats won't move anything you can show a board.

Tracking state is a platform problem, not a model problem

Of everything in the report, tracking state topping the blocker list is the most useful line for anyone running an engineering org.

‍

An agent that works for twenty minutes across six tool calls, two API failures, and a retry needs durable execution, idempotent side effects, and a record of what it actually did. Those are runtime properties. They belong to the platform your agents run on, not to whichever model you standardized on this quarter. Swapping models won't fix a workflow that loses its place halfway through and re-charges a customer on the retry.

‍

Security shows up the same way. 39.5% name it as the biggest barrier to letting agents run without a human watching, and the fix is a list of platform capabilities: a scoped, short-lived identity per agent per task, propagated end-user permissions, tool allowlists, full activity logging, and a rollback path. Autonomy gets earned by building those. It doesn't arrive with a better model.

The finding to put in front of your platform team

YouTube is the number one place engineers go when an agent misbehaves, ahead of GitHub and Stack Overflow. The report notes they'll check videos, Discord groups, chat groups, academic sources, GitHub, and X before they talk to someone at their own organization.

‍

Treat that as a maturity score. When there's no internal paved route for "my agent is stuck," people go outside for one. Every one of those searches is an engineer re-solving a problem three colleagues already solved, in a way nobody can audit, with a fix that lands in one repo and nowhere else. That's how you get four conventions for the same agent workflow inside one company, and it's a big part of why pilots don't spread.

What we'd measure instead

Seats issued and prompts run go up on their own. They're not evidence.

‍

Start with the delivery metrics from the DevOps Research and Assessment program: lead time for changes, deployment frequency, change failure rate, time to restore service. Add decision latency, meaning the calendar time a change spends waiting at the approve and accept gates. Then track cost per completed workflow rather than tokens consumed, since 79.8% of the survey says cost is already limiting scope and almost nobody can tie spend to an outcome.

‍

Three things have to advance together for any of this to hold: the platform that executes agents, the operating model that specifies what they build, and the business intent that defines what correct means. You're only as autonomous as the weakest of the three. A strong platform with weak specs automates ambiguity. Strong specs with a weak platform produce agents that know what to build and can't ship it safely. We wrote up the full set of failure modes in 10 reasons fintech agentic AI workflows stall, and the platform half of the argument in agentic delivery without a platform foundation is selling sand. The upstream version, where product decisions become the constraint, is in scaling product management for agentic delivery.

‍

If you want a read on your own three pillars that doesn't come from the fastest team in the building, our agentic engineering maturity assessment scores each one separately and tells you which is holding the other two back.

‍

Nothing in Temporal's data says the models are your constraint. Go check where the days went.

‍