Gemini 4 Carbon: Our next frontier model
deepmind.googleΒ·6d ago

TL;DR
CUGA, Hermes, and Deep Agents were tested with the same model on various tasks, revealing that the choice of harness significantly impacts performance, with CUGA excelling in enterprise environments. Task completion rates improved from 30% to 81% depending on the harness used.
β¦ Why It Matters
Evaluate your current agent harness to ensure it aligns with your enterprise requirements and improves task performance.
Key Takeaways
How It Works
CUGA operates through a compact loop of preparation, model calling, execution, and repetition, allowing it to scale from single agents to coordinated groups. It incorporates features like task decomposition and verification, which help catch mistakes before they affect outcomes.