AMD says AI agents now fix 75% of Radeon software issues

A bug report enters AMD’s Radeon Software eXperience, and an AI agent begins sorting the evidence: which requests are related, which code is likely involved, and which tests should follow. In June 2026, AMD says its agents automatically fixed 75 percent of reported RSX issues, up from 6 percent when the company started using them in October 2025.

RSX is the interface AMD uses to configure and monitor graphics-driver behavior. The agents do more than write code. They analyze problem reports, propose and implement changes, generate unit tests, identify integration and product-level tests, and assemble summaries for engineers to review before release. The system is designed to move through the software development lifecycle, not just autocomplete a function.

AMD also says the broader impact has exceeded its earlier target. After hoping for a 25 percent productivity boost over two or three years, the company reports a 30 percent overall gain after one year. Its main objective measure is accepted AI-generated source code: only code that passes review and testing and enters the final product is counted. That share has crossed 20 percent across the codebase, while some components exceed 80 percent.

The change AMD is pursuing next is structural. Instead of teaching one agent how an engineer normally solves a problem, the company wants swarms of agents to generate several approaches, validate correctness, measure performance, compare trade-offs and return ranked options. Engineers would define the issue, the desired outcome and the quality, performance and system constraints, then approve the result.

So what, concretely? For software teams, the immediate promise is less time spent triaging reports, writing repetitive tests and preparing release material. AMD’s RSX figures suggest that improvement can be measured in a real internal workflow, but they remain the company’s own results. The agents are already being used inside AMD; the broader swarm model is still under development, and the source does not establish independent results or deployment beyond the company’s engineering work.