AMD Says AI Now Writes 30% of Its Code. The Next Step Is Letting It Think for Itself
The chipmaker's software teams crossed a big productivity milestone. Now they are redesigning the entire development process around AI agents that solve problems without being told how.
Key points
- AMD achieved a 30 percent overall productivity boost from AI use in software development by mid-2025, beating its original 25 percent target set just one year earlier.
- AI now generates more than 20 percent of AMD's total production code, with some components exceeding 80 percent AI-written code.
- AMD's AI agents went from automatically fixing 6 percent of reported graphics software bugs in October 2025 to more than 75 percent by June 2026.
- The company's next goal is "agent swarms," groups of AI systems working in parallel to discover solutions independently, with engineers setting the goals rather than the steps.
A year ago, AMD told its engineers to expect AI to lift their productivity by roughly 25 percent over two to three years. They hit 30 percent in twelve months.
That number, published by IEEE Spectrum AI, comes with a specific definition. AMD counts only code that AI writes, passes all reviews, and ships in the final product. By that measure, more than one in five lines of production code at AMD now comes from an AI tool. In certain parts of the codebase, the figure exceeds 80 percent.
How did AMD's AI agents actually perform?
The clearest proof is in its graphics driver software. Starting in October 2025, AMD turned AI agents, software programs that can carry out multi-step tasks without human hand-holding at every turn, loose on bug reports in its Radeon Software eXperience (RSX) tool. RSX is the panel that lets graphics card owners adjust performance settings and monitor their hardware.
Out of the box, those agents fixed 6 percent of reported issues automatically. That is a modest start.
What happened next is the interesting part. AMD engineers did not retrain the underlying AI model. Instead they built a feedback loop, analysing every failure, refining the instructions given to agents, and letting the agents try multiple approaches until one passed the quality bar. Better AI models from outside vendors helped too. By June 2026, the automatic fix rate had climbed to more than 75 percent.
That is a twelve-fold improvement in eight months, without rebuilding the AI from scratch.
What is an "agent swarm" and why does AMD want one?
Right now, engineers teach AI agents to work the way the engineers themselves work. That is useful, but it caps the AI at human habits and human assumptions.
AMD's stated next step is a "swarm" model. Instead of one agent following detailed instructions, a group of agents would work in parallel, each trying a different approach to the same problem, then comparing results against clear success criteria before presenting ranked options to a human engineer for final approval.
Engineers would define what needs solving and what good looks like, not how to solve it. The agents figure out the path.
| Stage | Current approach | Swarm approach (planned) |
|---|---|---|
| Problem framing | Engineer writes detailed steps | Engineer sets goals and constraints |
| Solution finding | One agent follows instructions | Multiple agents explore in parallel |
| Evaluation | Engineer reviews output | Agents compare and rank options automatically |
| Approval | Engineer approves | Engineer still approves final choice |
AMD is already using multi-agent tools including Codex and Claude Code, the AI coding assistant built by Anthropic, alongside its own internal systems.
What does this mean for AMD's engineers?
AMD is explicit that headcount reduction is not the goal. The company says it wants engineers spending less time writing repetitive code and more time deciding what to build and why.
That framing is company policy for now, not a binding commitment. But the practical shift is real: as agents handle more implementation work, the human job tilts toward setting specifications, reviewing AI outputs, and catching mistakes the agent did not know it was making.
For engineers at any software company watching this, the skill that holds its value is judgment, knowing whether the AI's answer is actually good, not just technically passing.
Common questions
Does AI-generated code mean lower quality software?
Not by AMD's own measure. The company counts only AI code that clears every review and test before shipping, so the quality bar is the same as for human-written code. Whether that standard is sufficient is a separate question reviewers and regulators are still working through.
Will this kind of AI replace software engineers?
AMD says no, and frames its investment as giving engineers more capacity, not fewer jobs. The honest answer is that the role is changing: fewer hours on routine coding, more hours on oversight, design decisions, and validating what the AI produces.
