AMD: 75% of RSX Bugs Fixed by AI Agents by 2026

AMD reports that by June 2026, agents had automatically resolved 75% of issues with Radeon Software eXperience, compared to just 6% when the system was launched in October 2025. The company also claims a 30% increase in productivity thanks to AI and over 80% of code generated by models in certain components. It is preparing multi-agent workflows supported by a continuous learning loop, while massively training its teams and without announcing any layoffs.

By June 2026, 75% of RSX tickets automatically resolved

AMD cites the case of Radeon Software eXperience, an interface component for configuring and monitoring the behavior of graphics drivers. AI agents were deployed starting in October 2025 to diagnose and apply fixes to reported issues. At that time, off-the-shelf tools resolved only 6% of tickets. According to the company, the share of automatic corrections then progressed to reach 75% by June 2026. To achieve this, AMD states it established a learning loop, initially largely manual, aimed at analyzing agent failures and adjusting their objectives, rather than retraining the underlying models. Concurrently, model improvements and agent execution times reportedly enhanced efficiency, resulting in a claimed increase from 6% to over 75% of RSX issues resolved by this agentic loop.

A tracked indicator: 30% gain and up to 80% AI-generated code

AMD says it measures its progress through the percentage of source code generated by AI, counting only the code that has passed reviews and tests and is integrated into the product. According to this indicator, the 20% threshold was surpassed at the beginning of the year, with a projected trajectory towards 50% across the entire codebase and peaks exceeding 80% depending on the components. One year after its initial implementations, the company reports an overall productivity gain of 30% attributed to AI, surpassing the originally set goal of a 25% gain over two to three years. It recalls having launched in 2024 systems covering code generation, testing, bug analysis, and reviews, targeting 25% of production code generated by AI by 2027 and increasing automation of the SDLC.

Multi-agent workflows and a continuous learning loop

AMD describes an improvement mode still very centered on the engineer, who adjusts an agent's prompt and iterates. To move beyond this framework, the company argues that agents must learn from each other, reuse winning strategies, and progress collectively across projects and teams. It states that it is already extensively using multi-agent workflows through harnesses like Codex and Claude Code, while simultaneously developing internal multi-agent systems. AMD believes a continuous learning loop is necessary that reinjects errors and human interventions into future workflows, structured around clear and measurable objectives. Each cycle is presented as capturing new ideas, making AI engineering more efficient; ultimately, this self-reinforcing loop, rather than just the model, is expected to become a major driver of progress.

AI covers analysis, debugging, testing, and production deployment

AMD claims to have integrated agentic AI at all stages of its development cycle. For analysis and sorting, agents process reports, group similar requests, and highlight areas of code to modify. In debugging and generation, they interpret requests and make changes. For testing, they create unit tests and, if they pass, identify integration and product-level tests. Finally, for validations and releases, they produce an architecture summary, a review of changes, and a test results dossier for approval by engineers, then integrate the changes into the next release after receiving the green light.

Swarms of agents: problem defined by humans, solution sought by AI

AMD envisions swarms of collaborative agents capable of designing solutions without detailed prescriptions, with humans framing the problem, expected outcome, and quality, performance, and system constraints. In this framework, agents would generate, evaluate, and refine multiple approaches in parallel, validate corrections, measure performance, test trade-offs, and compare implementations against defined success criteria, before submitting ranked options with metrics and validations to engineers. The company presents this step as the greatest transformation to come, following an initial phase where AI primarily imitated human processes, against a backdrop of rapid model advancements and a reevaluation of the SDLC structure itself.

Teams and skills: empowerment rather than workforce reduction

AMD positions AI as a lever to increase productivity and quality and to redirect employees towards higher-value tasks. The company indicates it aims to equip its workforce rather than reduce headcount and states it is investing heavily in AI training across the organization. It wants every employee to master these tools confidently, responsibly, and effectively. AMD anticipates that as agent capabilities increase, engineers will spend less time on manual implementation and more on defining specifications, validating results, and making strategic decisions.