Production AI Agents Break Daily, But Trust Remains High

The survey. Temporal’s 2025 State of Development survey, drawn from 554 engineers and leaders actively building with AI agents, reveals a striking contradiction. 80.8% use agents daily, yet only 49.1% run them in production. The gap between usage and deployment is 30 points — and it’s filled with reliability problems. 41.1% of teams hit agent issues daily or more often; 16.4% see them every hour. Despite this, 85.5% trust agent output at least somewhat. Most teams trust a system that breaks every day.

The real bottleneck. When asked what limits productivity, tracking state across services was the top blocker (35.7%), followed by debugging (30.9%). Model capability didn’t make the top two. Cost is also a factor — 80% of respondents say cost limits how they use agents. The implication: production AI is no longer a model problem. It’s an infrastructure and state management problem. Teams are spending their time on recovery code and manual debugging, not on business logic.

The recovery gap. Only 13% of respondents are very confident they can observe and debug AI workflows at scale. 22% lose progress because of crashes. The gap between seeing a failure and recovering from it is a person reading traces at 2 a.m. and fixing state by hand. Observability alone doesn’t solve recovery — it just tells you what broke. Most teams lack the mechanics to resume work without losing the agent’s progress or re-spending LLM tokens.

The Temporal argument. The talk demonstrates a durable execution pattern that preserves workflow state across process crashes. A nine-step payment workflow — with an AI agent performing fraud investigation and a human approval step — recovers seamlessly after a worker is killed. Key effect: the agent’s tool calls run only once, avoiding retoken costs. The event history serves both as observability and as the source of truth for replay. The claim is that one platform can close all four identified gaps.

Production evidence. The demo is not theoretical. OpenAI Codex and Replit agents run on Temporal. The speaker notes that Replit’s engineering team found catastrophic errors in long-running agents to be a bad user experience, and durable execution addresses that. However, the evidence remains vendor-sourced — the survey was conducted by Temporal, and the speaker works there. Independent validation of the failure rates would strengthen the case.