A database locking issue in Zendesk's AI Agents service caused bot conversations to enter repeat loops between 14:40 and 17:00 UTC on July 14, 2026, with some instances persisting even after customer disengagement or human agent intervention. The root cause traced to a recent change in a cleanup job that increased query timeouts to manage a growing backlog. Under production load, the extended query duration held locks on critical database tables, blocking the live webhook processing path and triggering "lock wait timeout" errors. This prevented conversation state updates, forcing AI Agents to resend the same messages repeatedly. Zendesk identified and rolled back the problematic change within 37 minutes of the initial status update, with error rates dropping immediately and customer reports confirming resolution by 18:13 UTC.
The incident exposes a structural vulnerability in how background maintenance processes interact with live customer-facing systems when they share database infrastructure. For teams running AI Agents in production, the three-hour window represents genuine customer friction—repeated bot messages erode trust in automation and force manual escalation, undermining the efficiency gains these tools promise. The remediation roadmap (query optimisation, enhanced monitoring for lock-wait timeouts, and stricter rollout safeguards for database-impacting changes) suggests Zendesk recognises the gap between testing environments and production load, but the question for CX leaders is whether this incident reflects a broader pattern of AI agent reliability issues or an isolated database architecture problem. Teams already dependent on Zendesk's AI Agents should audit their incident response procedures for bot failure scenarios and consider whether their escalation workflows can absorb sudden automation failures without degrading customer experience.
The incident also highlights a less visible risk: the compounding effect of background job backlogs on system stability. As data volumes grow, maintenance processes designed for smaller datasets can become performance bottlenecks if not rearchitected. For organisations evaluating AI agent platforms, this serves as a reminder that vendor infrastructure maturity matters as much as feature capability—a platform's ability to handle scale without cascading failures directly impacts your support team's ability to deliver consistent service.
SummaryOn July 14, 2026, between 14:40 UTC and 17:00 UTC, some customers experienced AI Agents getting stuck in loops and sending repeated messages in bot conversations. In some cases, this continued even after an end user stopped responding or after a human agent replied. Timeline July 14, 2026 17: