Andrej Zdravkovic, a senior vice president at AMD, argues in a piece published on August 17, 2026 in IEEE Spectrum that software engineering will move from AI copilots mimicking human work to autonomous "agent swarms" that rewrite the software development lifecycle themselves. To back the claim, the company cites a 30 percent productivity boost from AI.
Key takeaways
- The piece is an opinion by Andrej Zdravkovic, an AMD SVP, in IEEE Spectrum (August 17, 2026) — not a product announcement.
- AMD reports a 30 percent productivity boost from AI, above its initial 25 percent target.
- AI-generated code crossed 20 percent in early 2026 and is heading toward 50 percent across the codebase, exceeding 80 percent in some components.
- RSX (Radeon Software eXperience) case study: an agentic loop resolved 6 percent of issues in October 2025 and more than 75 percent by June 2026.
- AMD uses agentic harnesses such as Codex and Claude Code in a multi-agent workflow.
Copilot versus swarm
Today's coding assistants are built to support a human at every step — suggesting a line, completing a function, fixing a bug. Zdravkovic argues that is also their ceiling: by mimicking human thinking, they inherit its limits. The alternative is collaborative groups of agents working in parallel that search for a solution on their own instead of replaying known steps.
The agents won't be enhancing each step of the SDLC—they will be rewriting the SDLC themselves.
Andrej Zdravkovic, senior vice president, AMD.
Proof from its own backyard: RSX
AMD's strongest argument is its own RSX component — the interface for configuring and monitoring graphics-driver behavior. The company says an agentic loop that closed a mere 6 percent of issues in October 2025 was resolving more than 75 percent by June 2026. That is an order-of-magnitude jump in under a year, though still on a narrow internal slice rather than the full development process. To build such workflows AMD reaches for ready-made agentic harnesses?Agentic harness: The runtime layer that gives a model tools, memory and an action loop — turning a bare LLM into an agent that can act on its own., including Codex and Claude Code.
| Metric | Value |
|---|---|
| AI productivity boost | 30% (initial target: 25%) |
| AI-generated code — whole codebase | heading toward 50% |
| AI-generated code — some components | over 80% |
The engineer's new role
If agents take over the actual writing of code, the engineer's center of gravity shifts from execution to specification and validation — precisely describing what should be built and checking whether the result is correct. That inverts today's setup, where implementation, not problem definition, eats most of the time.
Why it matters
The voice here is a chipmaker's executive, not an AI-tooling vendor — a signal that multi-agent coding is entering large engineering organizations, not just startups. The figures, though, are self-reported and cover narrow slices, so they are hard to compare with rivals. The value of the piece lies more in the direction than in hard proof that agent swarms work across a full process.
What's next?
- AMD points to AI-generated code heading toward 50 percent across the codebase — the next test is holding quality at that share beyond selected components.
- The RSX case (6 to more than 75 percent) is an internal case study — there is no independent third-party verification of the results.
Sources
- IEEE Spectrum — From AI Copilots to Agent Swarms
- Anthropic — Claude Code
