# UC Berkeley AI research developments

> Developing situation record from CLSTR: https://clstr.news/situations/uc-berkeley-ai-research-developments
> Updated: 2026-11-10T01:33:21.000Z. Sources: 4. Developments: 2.

Researchers at UC Berkeley have introduced several advancements in artificial intelligence. One development involves a multi-agent system called “Evolving Agent Teams” designed to optimize GPU kernels. This system uses a supervisor agent to monitor coding agents and dynamically restructure roles and coordination protocols to overcome performance plateaus. Testing showed the system achieved significant speedups on various hardware, including NVIDIA and AMD chips.

Further research has introduced “agentic robotics,” a method using AI agent teams to generate, test, and refine control programs within physics-based simulations. This approach aims to decrease the necessity for manual engineering and large imitation learning datasets. Additionally, researchers are advocating for the use of smaller, open-weight AI models hosted on local servers to ensure data control and the reproducibility of scientific results.

## Timeline

### 2026-11-10: UC Berkeley researchers develop agentic robotics and local AI models

UC Berkeley researchers have introduced agentic robotics to speed up robot programming and proposed using smaller, local AI models to improve data control and research reproducibility.

2 sources. https://clstr.news/cluster/uc-berkeley-researchers-develop-agentic-robotics-and-local-ai-models

### 2026-11-09: UC Berkeley researchers develop evolving AI agent teams for GPU optimization

UC Berkeley researchers developed an AI multi-agent system that dynamically reorganizes its team structure to optimize GPU kernels, outperforming existing production code in several tests.

2 sources. https://clstr.news/cluster/uc-berkeley-researchers-develop-evolving-ai-agent-teams-for-gpu-optimization

---
Cite as: UC Berkeley AI research developments. CLSTR, https://clstr.news/situations/uc-berkeley-ai-research-developments
