UC Berkeley Team Builds Self-Optimizing AI Agents

Researchers report a multi-agent system that restructures itself to write GPU kernels that outperform expert-tuned production code on most tested workloads.

Updated on Nov. 9, 2026 in Artificial Intelligence

UC Berkeley Team Builds Self-Optimizing AI Agents

UC Berkeley researchers published Evolving Agent Teams, a multi-agent system for GPU kernel optimization in which a supervisor agent restructures a team of coding agents when progress stalls. The paper reports that the approach outperformed the production FlashInfer kernel on 36 of 47 workloads.

Why it matters

The work suggests that AI systems coordinating multiple agents can automate low-level performance engineering that has traditionally required specialized human experts. Because GPU kernel efficiency directly affects the cost of running large models, gains in this area bear on the economics of AI computing.

The paper reports a 1.21x geometric mean speedup over the production FlashInfer kernel, with evolving teams outperforming the baseline on 36 of 47 workloads and reaching 3.01x over vLLM on AMD's MI350X and 3.44x over FlashKDA on NVIDIA's B200. Whether these gains hold outside the tested workloads and configurations has not been established.

The players

UC Berkeley

The University of California, Berkeley is a public research university whose computer science researchers produced the work.

Berkeley Sky Computing Lab

The Sky Computing Lab is a research group at UC Berkeley focused on cloud and distributed computing systems, which developed the project.

The details

The system stores learning documentation, a history of team evolution, and agent artifacts in a persistent shared file system. Its supervisor agent diagnoses when work has plateaued and changes agent roles and coordination, with the team exploring over 100 configurations on a single task.

- October 27, 2026: The Evolving Agent Teams paper was published as arXiv:2610.36848.

Tech Race

The research measures its results against FlashInfer, the production GPU kernel library that many AI systems rely on, and compares them with single-agent and fixed-role setups. The comparison positions multi-agent coordination as a competing method for automating kernel engineering.

For people who use AI services, faster GPU kernels can translate into lower compute costs and quicker responses over time. The benefits depend on whether these techniques are adopted in production libraries that cloud providers and developers rely on.

The takeaway

Automated agent coordination may become a practical tool for performance engineering, but readers should watch for independent replication before assuming the reported gains apply broadly. Developers working with GPU workloads can track the paper's methods as a reference point for multi-agent optimization approaches.