Temporal Survey: AI Agents Trusted, Yet Break Every Day

What Temporal's Survey of 554 Engineers Says About AI Agents in Production

Shy Ruparel is a developer advocate at Temporal, the durable execution platform. In this AI Engineer talk he presents Temporal's second annual State of Development survey of 554 engineers and engineering leaders who actively use AI agents. Nearly half have 11 or more years in the industry, two thirds are in the US and a third in the UK and EMEA. Ruparel organizes the results around four gaps between how teams use agents and how reliably those agents run, then demos an agent-assisted payout workflow surviving two worker crashes.

Gap 1, usage versus production: Daily agent use rose from 47.3% in the 2025 survey to 80.8% this year. But only 49.1% say they run agents in production or that agents are core to how they ship. The other half are still experimenting or using a single agent, so roughly 30 points separate daily use from production.

Gap 2, trust versus reality: 85.5% of respondents trust agent output at least somewhat, and a quarter trust it almost completely. Yet 41.1% hit agent issues daily or more often, and 16.4% hit them every hour. Teams that call themselves successful were 1.7 times more likely to hit issues continuously, and the number of agents a team runs doesn't correlate with how often issues occur. Ruparel's guess is that successful teams look harder. His conclusion: "failure isn't going away. You have to account for it. You have to plan for it." Separately, 40% of teams say security is what blocks them from autonomous agents.

Gap 3, prototype versus enterprise deployment: 51.3% can take an AI prototype to production-ready in hours or less. At companies with 5,000 or more employees, about one in five say it takes months, against about one in 20 everywhere else. Ruparel attributes the difference to what banks, insurers and payment companies require: approvals, audit trails and systems that never lose a payment.

Gap 4, seeing versus recovering: The top productivity blockers are not the models. 35.7% name tracking state across services, 30.9% name isolating and debugging issues, and 22.2% lose progress and redo work after crashes and restarts. Only 13% are very confident they can observe and debug AI workflows at scale. The most common place engineers go for help when an agent breaks is YouTube, which Ruparel calls what debugging looks like "when your system can't tell you what happened." Eight in ten say cost limits how they use agents, which makes rerunning an agent from scratch after every crash expensive.

The payout demo: Ruparel runs a nine-stage payout workflow on a local open-weight model: validate the request, reserve funds, fetch an FX quote, then an agent investigates a suspicious transfer, looks up the customer and drafts a brief for a human approver. He kills the worker mid-run and approves the payment while no worker is running. In a typical stack, the funds would stay reserved at the bank, the brief would be lost and the tokens would have to be spent again. With durable execution, the approval is recorded as a signal in the event history and a new worker resumes from the last completed step. Each AI step ran exactly once. The workflow owns the money and the state, while the agent only advises: "if the model had failed, the payout would have continued without the brief." Retries are a declared policy, not a loop, and bad model output is treated as an ordinary activity failure. Ruparel names OpenAI's Codex and Replit's agent as running on Temporal in production.

5 Takeaways from Temporal's State of Development Survey

• Daily use is not production - 80.8% use agents daily but only 49.1% run them in production
• Trust doesn't mean reliability - 85.5% trust agent output, yet 41.1% see agent issues daily or more often
• Large companies ship slower - About one in five companies with 5,000+ employees need months to take an agent to production
• State is the bottleneck, not the model - Tracking state (35.7%) and debugging (30.9%) top the list of blockers
• Keep agents advisory inside deterministic workflows - Let the workflow own money and state, and let the agent's output be optional so a model outage degrades nothing

What Temporal's Data Means for Running AI Agents in Production

The survey says agent failure is routine even for teams that trust their agents, so the design goal is making failures cheap rather than rare. For organizations putting agents into long-running, money-moving processes, that means recording every step, approval and model call so a crash resumes instead of restarting, and keeping a human in the loop for the irreversible step. The vendor framing is Temporal's, but the reliability numbers apply to any agent stack.