AMD Says AI Agents Now Auto-Fix 75% of Radeon Software Bugs

AMD reports that AI agents now automatically resolve more than 75 percent of reported issues in its Radeon Software eXperience (RSX) user interface, up from just 6 percent when the effort began in October 2025. The figure anchors a broader shift the company describes in a recent essay: AI has moved from generating lines of code to reshaping the software development lifecycle (SDLC) itself.

The stakes are straightforward. AMD set a goal in 2024 of 25 percent AI-generated production code by 2027, alongside a hoped-for 25 percent productivity boost over two to three years. The company says it surpassed that productivity target in a single year, hitting a 30 percent overall boost through AI.

Measuring developer productivity is notoriously difficult — AMD cites ongoing research on the problem — so the company tracks one hard metric: the percentage of source code generated by AI that passes all reviews and testing and ships in the final product. By that measure, AMD crossed 20 percent at the beginning of this year and is now pushing toward 50 percent across its entire codebase. In some software components, more than 80 percent of shipped code is AI-generated.

Agentic AI now touches every phase of the lifecycle. Agents analyze problem reports, group similar requests, and flag code likely to need modification. They debug, implement code changes, generate unit tests, and escalate to integration and product-level tests when units pass. At the approval stage, agents prepare architecture summaries, code change reviews, and full test results for engineers to sign off, then integrate approved changes into the next release.

From copilots to swarms

AMD argues the biggest transformation is still ahead. Today's agents are built "in their own image," as the company puts it: engineers teach AI what they know about a system and how they would fix an issue. That lets engineers create multiple AI versions of themselves working in parallel, but it remains, in AMD's words, "constrained by human thinking and human-defined approaches."

The next step is collaborative agent swarms that work independently. Instead of detailed instructions, engineers would define the problem, the desired outcome, and the quality, performance, and system constraints. Swarms would then generate, evaluate, and refine multiple solutions in parallel, validate correctness, measure performance, test trade-offs, and present ranked options with validation results for human approval. "The agents won't be enhancing each step of the SDLC—they will be rewriting the SDLC themselves," AMD writes.

The RSX debugging effort shows how fast agent performance improves with the right feedback loop. Out-of-the-box AI tools resolved only 6 percent of issues. Rather than retraining the underlying models, AMD refined the objectives given to agents, letting them iteratively explore approaches and converge on better solutions while model and runtime advances compounded the gains. The company credits this continuous learning loop — feeding errors and human interventions back into future workflows — as the mechanism that will drive AI progress, not just raw model improvements.

AMD says it already uses multi-agent workflows extensively through agentic harnesses such as Codex and Claude Code, while building internal multi-agent systems for the next generation of AI-driven engineering.

The human role

AMD frames the shift as augmentation, not replacement. "Our goal is to empower our workforce with AI, not to reduce headcount," the company states, noting heavy investment in AI education and training across the company. As agents improve, engineers will spend less time implementing solutions and more time defining specifications, validating outcomes, and making strategic decisions.

If AMD's trajectory holds — from 6 percent to 75 percent automated bug resolution in under a year — the open question for the industry is how quickly swarm-based development moves from internal experiments like RSX to the broader engineering organization.