AgentKernelArena documentation

2026-07-14

1 min read time

Applies to Linux

AgentKernelArena is a controlled experimentation platform for developing AI agents on real GPU kernel optimization tasks. It enables reproducible A/B tests across models, prompts, tools, and agent policies, and produces objective compilation, correctness, and performance signals that external agent-RL systems can use as rewards.

The AgentKernelArena source code is hosted in the AMD-AGI/AgentKernelArena GitHub repository.

Install

• Install AgentKernelArena

How to

• Run an experiment

• Run tasks in parallel across multiple GPUs

• Configure agents and models

• Add a task

• Validate tasks

• Visualize and compare runs

Examples

• Examples

Reference

• Configuration and API reference

• Performance measurement methodology

To contribute to the documentation, see the AgentKernelArena GitHub repository.

AgentKernelArena is released under the Apache 2.0 license. For details, see the License page.