Why AI agent failures aren't always about the model

As AI agents move into production environments, they're introducing a new debugging challenge: when they fail, the culprit often isn't the model itself. Unlike traditional software that tends to surface clear errors, an autonomous agent can run successfully while heading in the wrong direction, masking the real cause of failure deep in its decision-making process.

The observability gap

Nvidia VP of Product Adel el Hallak recently explained to The New Stack that conventional debugging won't cut it for agent-based systems. Checking logs, inputs, and outputs isn't enough when an agent has autonomously chosen its own tools and course corrections along the way. The problem is fundamentally different from diagnosing a crashed service or a failed API call.

Even high-performing coding agents stumble more than 60% of the time on real-world tasks. But knowing an agent failed doesn't answer the crucial follow-up: why did it fail? Diagnosing the root cause requires understanding its entire reasoning chain—which tools it selected, where it got stuck, and when it pivoted to a different approach.

A three-layer stack

Nvidia breaks the agent architecture into distinct layers: the model provides reasoning capability, the harness orchestrates execution, and the runtime governs security and policy enforcement. When something goes wrong, the finger shouldn't automatically point at the model.

Recent Nvidia research found that swapping out the harness while keeping the model constant can meaningfully improve performance. This means a poorly tuned orchestration layer can drag down an otherwise capable model. El Hallak notes that models vary in their behavior—some are more verbose than others—so pairing them with appropriately tuned harnesses is increasingly critical to success.

Nvidia's OpenShell runtime is positioned as the non-negotiable foundation, capturing visibility into agent execution while managing sandboxing and policy enforcement. This visibility is what makes it possible to replay an agent's steps and identify where execution diverged from the intended path.

Shared vulnerability disclosure for agents

Recognizing that agent failures pose industry-wide risks, Nvidia is backing the Secure Agent Findings Exchange (SAFE), an initiative supported by roughly 140 companies. The goal mirrors vulnerability disclosure in traditional software: when one organization discovers how an agent can fail, everyone benefits from the fix.

Traditional software has long-standing infrastructure for reporting and patching bugs across the ecosystem. SAFE attempts to establish similar mechanisms for agent failures. This becomes increasingly important as specialized agents handle critical workflows—failures in a security agent patching systems, for instance, could originate in the model, the execution path, or the tools themselves, making them harder to isolate.