What happened
Nvidia researchers published new findings on Friday showing that the software wrapper around an AI model — the “harness” — is critical for complex tasks. As reported by TechCrunch, their custom harness allowed Claude Opus 5 to achieve a 100% score on the ARC-AGI-3 interactive reasoning benchmark. Without this specialised scaffolding, the same model scored just 30%, which was still the highest result among all base models tested.
How the room's reading it
The research adds to a growing consensus among practitioners. The chatter on X and in dev forums points to a shift in focus from model choice to the design of the harness — the system that handles memory, tools, and feedback loops. This echoes recent research from Databricks, whose CEO noted that a poor harness can easily double an application's cost. Some see Nvidia's move as a direct challenge to OpenAI's more closed approach, framing open agent stacks as the path to more control and better, more secure performance for builders.
Sailfish's take
We see this as a huge win for builders. It confirms something we've seen in our own work — the model is just the engine, but the harness is the vehicle. You don't need a frontier model to ship a robust agent. This research gives teams permission to stop chasing the latest API and focus on clever system design instead. A well-built harness on a cheaper, faster model can outperform a lazily-wrapped frontier model for many real-world tasks. If you're still just wrapping a single API call, this is the week to start thinking about your agent's memory, tools, and supervision loop.