AMD says AI already lifted software productivity 30 percent, eyes agent swarms next

Andrej Zdravkovic, a senior vice president at AMD with more than thirty years in software development and leadership, wrote in IEEE Spectrum that AMD's use of AI across the software development lifecycle has now surpassed the company's own target. A year earlier AMD had been hoping for a 25 percent productivity boost from AI over two to three years; the company says it has already achieved a 30 percent overall productivity boost.

AMD's other tracked metric is the share of source code generated by AI that passes all reviews and testing and ships in the final product. The company set a goal of 25 percent AI-generated production code by 2027. It crossed 20 percent at the start of this year and is now progressing toward 50 percent across the entire codebase; in some software components, Zdravkovic says, more than 80 percent of the code is now AI-generated.

Agentic AI is already applied at every stage of AMD's development lifecycle: agents triage and group incoming problem reports and flag likely code to change, analyze bug requests and implement fixes, generate unit tests and then identify needed integration and product-level tests, and finally prepare architecture summaries, code-change reviews and full test results for engineers to approve before release.

Zdravkovic argues this current generation of agents is still built "in engineers' own image": staff teach the AI how they personally would fix an issue, so the agents remain constrained by human-defined approaches even as they scale individual productivity. He says the next transformation will be collaborative swarms of AI agents that independently generate, evaluate and refine multiple competing solutions in parallel, validating correctness, measuring performance and comparing implementations against success criteria set by engineers, who would then review ranked options rather than write the code themselves. AMD already runs multi-agent workflows through agentic harnesses including Codex and Claude Code, alongside internal multi-agent systems it is developing for this purpose, though the article gives no detail on how the two approaches are used or in what proportion.

As a case study, Zdravkovic points to AMD's Radeon Software eXperience (RSX), the user interface for configuring and monitoring graphics driver behavior. AMD began using AI agents to automatically debug and fix reported RSX issues in October 2025. Out-of-the-box AI tools initially resolved only 6 percent of issues. By building a learning loop that fed failures back into refined objectives for the agents, rather than retraining the underlying models, and by riding improvements in models and agent runtimes, AMD raised that resolution rate to more than 75 percent by June 2026.

Zdravkovic frames the company's overall approach as reallocating engineers toward higher-value work rather than cutting headcount, and says AMD is investing in AI education so its workforce can use these tools confidently. The article names no other AMD staff or executives, gives no team size or headcount figures, and does not specify which software components reached the 80 percent AI-generated-code mark.

Key facts

AMD says AI has driven a 30 percent overall productivity boost in software development, beating a 25 percent target it had set a year earlier for a two to three year horizon.
By AMD's metric of AI-generated code that passes review and testing and ships, the company crossed 20 percent at the start of this year, is heading toward 50 percent across its whole codebase, and already exceeds 80 percent in some components, against a 2027 goal of 25 percent.
AMD senior vice president Andrej Zdravkovic argues today's agents are still shaped by individual engineers' habits, and that the next step is collaborative agent swarms that independently generate and rank competing solutions for engineers to approve.
In AMD's Radeon Software eXperience (RSX), AI agents raised the automatic issue-resolution rate from 6 percent when the effort began in October 2025 to more than 75 percent by June 2026, through a feedback loop that refined agent objectives rather than retraining models.
AMD already uses agentic harnesses including Codex and Claude Code alongside internal multi-agent systems, and frames the initiative as reallocating engineers to higher-value work rather than cutting headcount.

Why it matters

This is a rare, numbers-attached public account of agentic AI in production software engineering from a major chip and computing company, not an AI vendor. It documents a specific trajectory: from AI as a coding assistant measured by a productivity percentage, to AI-generated code approaching half of an entire codebase, to a stated ambition of autonomous agent swarms that would rewrite the software development lifecycle rather than just accelerate steps within it.

Who it affects

AMD's own software engineers, whose day-to-day workflow AMD says is already built around agent triage, debugging, test generation and release review. More broadly, it is a data point for other large engineering organizations weighing how far to push agentic coding tools, and for anyone tracking how the industry talks about productivity targets and AI-generated code share as concrete, trackable metrics rather than vague claims.

How to use it

AMD names Codex and Claude Code as agentic harnesses it already runs multi-agent workflows through, alongside internal systems it is building; the article gives no pricing, licensing or deployment detail, so there is nothing actionable here beyond the fact of that usage. The RSX case, where AMD moved from 6 percent to over 75 percent automatic issue resolution by refining agent objectives rather than retraining models, is offered as a template for building a feedback loop around agentic tools.

How solid is it

The figures come directly from an AMD senior vice president writing in his own voice for IEEE Spectrum, not from an independent audit. AMD does describe its methodology for the headline metric with some rigor, counting only AI-generated code that passes all reviews and testing and ships in the final product, which limits inflation from throwaway suggestions. Even so, every number in the piece is self-reported by the company being measured, and IEEE Spectrum is publishing an opinion-style byline rather than reporting it independently.

Risks and caveats

The article does not say which software components reached the over-80-percent AI-generated-code figure, nor how many engineers or teams are involved. The mechanism by which future agent swarms would "learn from one another" and "reuse successful strategies" is stated as a goal, not described as a working system. AMD's assurance that the effort aims to reallocate engineers rather than reduce headcount is a company statement about intent, not a verifiable outcome, and the vision of engineer-guided but otherwise autonomous agent swarms remains aspirational rather than something AMD reports having built.

“We believe the next major transformation in software engineering will occur when collaborative AI agent swarms can independently identify and develop solutions, guided by humans on what to solve rather than constrained by human assumptions about how the job should be done.”

— Andrej Zdravkovic, senior vice president at AMD