From AI Copilots to Agent Swarms
IEEE Spectrum Andrej Zdravkovic
AMD says AI now writes 30% of its code, up from a 25% target. The bigger bet: swarms of agents that don’t just copy engineers, but find fixes on their own.
Based on reporting by IEEE Spectrum, Andrej Zdravkovic — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
AMD says its use of AI in software work has moved faster than it expected. Last year, the company was aiming for a 25 percent productivity boost over two or three years. One year later, it says it has already hit a 30 percent overall boost, and that the real shift is still coming.
The current story is less about flashy code demos than about measurement. AMD tracks one hard number: the share of source code generated by AI and shipped after passing review and testing. By that yardstick, it says it crossed 20 percent at the start of this year and is now moving toward 50 percent across the whole codebase. In some parts of the software, it says, AI already generates more than 80 percent of the code.
The company is also pushing AI into more of the software development lifecycle. It says agents now help with triage, bug analysis, code generation, test creation, and release prep. In practice, that means reading problem reports, grouping similar requests, suggesting code changes, generating unit tests, and assembling the summaries engineers need before approval. AMD has been using multi-agent workflows through tools such as Codex and Claude Code while building its own internal systems too.
The example AMD leans on is Radeon Software eXperience, or RSX, a user interface layer for graphics driver settings and monitoring. In October 2025, the company started using AI agents to debug and fix reported RSX issues. Off-the-shelf tools only solved 6 percent of them. By June 2026, AMD says the rate had climbed to 75 percent after it built a learning loop around the failures and kept refining the goals given to the agents instead of retraining the base models.
That leads to the bigger argument in the piece: today’s agents still mirror human habits, but tomorrow’s should be able to generate and compare solutions in parallel, then hand engineers ranked options with validation data attached. AMD says that future will change the SDLC itself, not just speed up parts of it. It also says human engineers will spend less time typing fixes and more time defining what good looks like, checking outcomes, and making the decisions that matter most.
My take — AI-written commentary, not fact-checked reporting
AMD is making the right bet here: the real value isn’t a chatty code buddy, it’s boring systems that churn through fixes and tests without needing a meeting first. The industry loves to sell AI as a genius intern; the smarter move is to treat it like a factory line that still needs a human at the end. That’s less glamorous, and far more useful.
Read more about this at: IEEE Spectrum