Temporal survey: 80.8% of developers use AI agents daily, only 49% ship to production

Daily agent use is now the norm

Temporal's State of AI Agent Development 2026 survey, surfaced this week in a dev.to analysis, records a steep adoption curve: 80.8% of developers now say they use AI agents daily or more, up from 47.3% a year earlier. The survey covered 554 engineers, and 21.8% described their usage as continuous. Daily interaction with agents has effectively become the default mode of developer work.

Production tells a different story. Only 49.1% of respondents run agents in production, leaving a 31-point gap between everyday use and deployment. Compute is the other brake: 79.8% said compute costs meaningfully limit their progress with agents.

The verification bottleneck

The dev.to write-up argues the gap is not a timing problem but a verification problem: agents generate code and decisions faster than humans can check them, and unchecked output does not get shipped.

Survey data from Sonar backs that reading. According to Sonar, 96% of developers do not fully trust AI-generated code, only 48% always verify it before committing, and 38% find reviewing AI code more effortful than reviewing code written by people. Sonar's Tariq Shaukat, speaking at the AI Engineer World's Fair, said lines of code have risen dramatically while releases have grown only about 30%, and that human reviewers agree with incorrect AI output four times out of five.

Anthropic's Claude Code team published the sharpest internal numbers on the problem. There, changes requiring review climbed from roughly 2,500 to more than 30,000 per day within four months. Rather than slowing generation, the team moved verification into the build process itself, running checks while the agent works instead of afterwards. The reported result was twice as many changes shipped with roughly 40% fewer reverts.

The harness argument

At LangChain's Interrupt London event, Harrison Chase argued that defensible value now sits in the harness — the system around the model rather than the model itself. "The harness is the moat," he said, laying out five components: evals that test agent behavior against a team's own domain, persistent memory, reusable skills, governance such as audit trails and spending limits, and fine-tuning loops that feed production data back into the system. On the Sequoia podcast, Chase added that the companies best at this kind of engineering are coding companies rather than AI labs.

The dev.to analysis points to GitHub's trending page as corroborating evidence: on October 30, 12 of the 15 trending repositories were agent infrastructure, spanning memory, skills, quality gates, token efficiency and execution environments, with a combined star count above 1.29 million.

Trust in autonomous action

Correctness is one half of the production problem; autonomous action is the other. At the AI Engineer conference, SoFi's Ishween Kaur demonstrated six attacks on x402, an emerging standard for agent-to-agent payments, in which agents were induced to pay for nothing. Her framing — how an agent decides whether a five-cent payment is legitimate — is a governance question rather than a model-capability question, which is why spending limits, approval gates and audit trails appear in Chase's component list.

Skepticism is visible in developer communities too. A Hacker News thread titled "AI isn't replacing SREs, it's deskilling them" drew 466 points and 341 comments, with a top comment comparing mandated manual flying time for pilots to human-in-the-loop checkpoints for production agents.

Why it matters

The Temporal numbers are a benchmark: daily agent use among developers has moved from a minority position to the mainstream in a single year, and nearly half of developers already operate agents in production. But the 31-point gap between those two figures identifies where the next phase of work sits. The evidence assembled in the dev.to analysis — Sonar's trust data, Anthropic's review-volume spike, the attacks on agent payment protocols — consistently points away from model capability and toward verification, memory, skills and governance as the deciding factors in whether an agent ships. For engineering teams, the practical takeaway is that adoption is no longer the differentiator; the surrounding infrastructure is. For vendors and open-source maintainers, the wave of harness-oriented projects on GitHub suggests the market has already reached that conclusion.