Your AI agent failed. The model might not be the problem.
The New Stack Amanda Caswell
AI agents can fail without crashing, so the model may not be the thing to blame. Nvidia says you need to debug the runtime, harness and tools too.
Based on reporting by The New Stack, Amanda Caswell — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
AI agents are getting harder to diagnose just as people start putting them into production. Unlike older software, they can keep moving after a bad decision, switch tools on their own and drift off course without throwing an obvious error.
That is the problem Nvidia’s Adel el Hallak is pushing on. In an interview with The New Stack, he said developers need far more visibility into how an agent got to an answer, not just the logs and the final output. What did it reason through? Which tools did it touch? Where did it stall? Sometimes the only way to know is to replay the run and watch it go sideways.
Nvidia is also backing SAFE, the Secure Agent Findings Exchange, an industry effort supported by roughly 140 companies. The idea is to build shared infrastructure for reporting agent failures, borrowing from the way traditional software shares vulnerability disclosures. The logic is simple: if one team finds a nasty failure mode, the rest shouldn’t have to rediscover it the hard way.
Nvidia’s view is that the runtime is where much of this needs to be captured. Its OpenShell agent runtime, part of the NemoClaw platform, handles sandboxing and policy enforcement while exposing what the agent is doing. El Hallak called OpenShell the one non-negotiable piece across Nvidia’s reference architectures, even while saying the harness and the model can vary.
That distinction matters because Nvidia’s own research suggests the model is only part of the story. Its NOAH work showed that changing the harness while keeping the model fixed can improve performance. And in some cases, a bad harness can hold back a decent model. On top of that, OpenAI has said monitoring can add about 20% to inference compute for its most capable persistent agents, which tells you why a lot of teams would rather blame the model and move on.
The bigger shift here is that AI safety is starting to look less like a model leaderboard problem and more like systems engineering. Bugs in agents may come from the model, the tools, the runtime or the way they all interact. If that sounds messy, that’s because it is. The old habit of blaming the model first is convenient, but it’s also lazy.
My take — AI-written commentary, not fact-checked reporting
The industry loves to point at the model when an agent face-plants, because that keeps the story tidy. But the real failure is usually the stack around it, and everyone pretending otherwise is just doing software theater with a bigger budget. Shared failure reporting is the boring, necessary bit that might save a lot of future embarrassment.
Read more about this at: The New Stack