AI Agent Monitoring Emerges as Critical Infrastructure Need

The shift to production AI

As artificial intelligence systems transition from experimental deployments to production environments, technology leaders are confronting a new operational challenge: AI agents fail differently than traditional software, and those failures are harder to detect.

Jenn Tejada, Executive Chair of PagerDuty and former CEO who led the company through its 2019 IPO, draws on experience spanning multiple technology cycles to frame the current moment. Having worked through both the late-1990s internet boom and the cloud computing revolution, she sees AI's production phase creating unprecedented infrastructure demands—with hyperscaler capital expenditure projected to reach $725 billion in 2026, nearly double the previous year's spending, according to BNP Paribas estimates.

This infrastructure buildout extends beyond data centers into workforce development, exemplified by Meta's data center expansion paired with its America's Workforce Academy training initiative. The scale of investment is creating opportunities across the technology ecosystem, from experienced engineers to entry-level workers.

New failure modes require new detection

The operational challenge comes from AI's distinct failure characteristics. When traditional software breaks, it typically stops functioning. AI model drift operates differently—the system continues executing, but its outputs gradually degrade in ways that compound before becoming visible.

"When AI drifts, it's actually harder to see, and you don't see until it's executed that drift in a number of ways, and now it's evolved into multiple failures," Tejada explained in an interview first reported by Forbes contributor Martine Paris.

This detection gap makes early warning systems essential. AIOps platforms are evolving to monitor AI models alongside traditional digital infrastructure, providing what Tejada describes as an "unbiased opinion" of agent performance and enabling human intervention before minor issues cascade.

The complexity extends to organizational structure. Tejada anticipates the rise of "one and two pizza teams"—Amazon's term for groups small enough to feed with one or two pizzas—empowered by AI to tackle problems previously requiring larger engineering organizations. But smaller teams managing more complex systems increase the surface area for potential failures.

Why it matters

The shift from experimental AI to production deployment at scale introduces systemic risks that existing monitoring infrastructure wasn't designed to catch. Unlike traditional software outages that are binary and immediate, AI model drift degrades gradually and can propagate through interconnected systems before becoming apparent. For enterprises deploying autonomous agents in critical workflows—from customer service to supply chain management—the inability to detect drift early could mean discovering failures only after significant business impact. This creates a new category of operational requirement: continuous AI model monitoring with human oversight capabilities, distinct from traditional application performance management.

Learning from minor failures

Tejada frames the challenge within a broader perspective on innovation. The increased system complexity and new failure modes are trade-offs for advances spanning healthcare to transportation. The operational goal isn't eliminating all failures—an impossibility in complex systems—but rather avoiding catastrophic incidents while extracting lessons from smaller breakdowns.

This approach acknowledges that as AI agents gain autonomy, the monitoring systems watching them become as critical as the agents themselves. Whether a system failure originates from human error or agent malfunction matters less than having infrastructure capable of detecting and responding to problems before they escalate.

These details were first reported by Martine Paris for Forbes.