A tiny AI agent just beat bigger rivals — because of how it was built, not its size
New research from a London startup and from Nvidia both suggest the software wrapped around an AI model now matters more than the model itself.
What happened: London startup Inherent, founded by ex-Google DeepMind staff, says its AI agent Faraday beat Anthropic's Claude Opus 4.8 and OpenAI's GPT-5.5 at independently reproducing findings from published science papers, using a much smaller 27-billion-parameter model. Separately, Nvidia published research showing that swapping the 'harness' (the software wrapper handling memory, tools and oversight) around Claude Opus 5 pushed its score on the tough ARC-AGI-3 game benchmark from 30% to a perfect 100%, without touching the model itself.
Why it matters: For two years, the AI industry's main pitch has been: bigger, newer model, better results. Both stories complicate that story. Inherent got frontier-beating results from a fraction of the size by focusing on training method and 'research taste' rather than raw scale. Nvidia showed the same model can go from mediocre to flawless purely by changing the scaffolding around it. That reshuffles where the real value, and the real competition, in AI now sits.
How it works, plainly: A 'harness' is everything wrapped around a raw model: memory management, tool access, and rules for acting on its own over long tasks. Nvidia's version added a 'supervisor' agent that nudges the main agent when it stalls or wanders off track, like a manager checking in. Inherent trained Faraday using reinforcement learning, rewarding good research instincts rather than dictating rules, and had it borrow OpenAI's coding tool rather than build its own, mirroring how human scientists rely on existing tools.
The rollout: Inherent, which raised $50 million and has 12 staff, plans to grow to roughly 20-25 people by year's end and is aiming beyond replication toward agents that can generate new scientific findings. Nvidia's harness, called AVO, isn't a product but adds to open-source tooling under its Nemo brand. OpenAI and Databricks have separately found similar harness effects on scores and costs, suggesting this is becoming an industry-wide pattern rather than a one-off result.