AI is no longer limited to suggesting lines of code at AMD. The company says AI has delivered a 30 percent overall productivity boost, while more than 80 percent of the code in some software components now comes from AI.
That progress began with a more modest goal. AMD aimed for a 25 percent productivity boost from AI use over two or three years, but the company says it has already surpassed that target. The results point to a shift from isolated coding assistance toward AI-supported software development across major engineering workflows.
AMD Surpasses Its AI Productivity Target
On August 17, 2026, AMD reported that AI had produced a 30 percent overall productivity boost. Last year, AMD hoped to reach a 25 percent productivity boost over two or three years, setting a clear measure for its software development efforts.
“We have surpassed our productivity target, achieving a 30 percent overall productivity boost through AI,” AMD said.
The company also reported that AI now generates more than 80 percent of the code in some software components. That figure shows how deeply code generation has entered parts of AMD’s development process, moving beyond small suggestions or isolated experiments.
Yet generating code is only one part of software engineering. Code must also be tested, debugged, and adapted when users report problems. AMD’s work with AI agents in Radeon Software eXperience, or RSX, focuses on that larger challenge.
RSX AI Agents Improve From 6 Percent to 75 Percent
In October 2025, AMD began using AI agents to automatically debug and fix reported RSX issues. The first results were limited: the agents resolved only 6 percent of issues.
AMD described the early performance in direct terms: “Out-of-the-box AI tools delivered limited results, resolving only 6% of issues.”
That starting point created a demanding test for AI agents. Instead of measuring whether a system could produce code in a controlled task, AMD tracked whether agents could handle reported software issues and deliver fixes.
By June 2026, the percentage of RSX issues fixed automatically by AI agents had reached 75 percent. The change from 6 percent initially to 75 percent in June 2026 marks the clearest measure of AMD’s progress with autonomous software debugging and repair.
The RSX results also add a new dimension to AMD’s 30 percent productivity figure. Code generation can accelerate development, while AI agents can address reported issues. Together, these efforts cover more stages of software work.
MLPerf Client Expands Into Agentic Software Engineering
On August 18, 2026, MLCommons announced MLPerf Client v2.0, extending the MLPerf industry-standard benchmark for evaluating AI performance into new areas. The release includes new Image Generation and Agentic AI categories, with metrics for both responsiveness and throughput.
MLPerf Client v2.0 features Flux.2 klein 4B as an experimental test in Image Generation. Its Agentic AI category benchmarks performance through two scenarios: Software Engineering, also called SWE Agent, and Data Analyst Agent.
Those scenarios bring software development into a benchmark designed to measure how AI performs in practical client workloads. The SWE Agent scenario connects directly with the type of coding and issue-resolution work AMD has been developing, while the Data Analyst Agent scenario adds another agentic workflow to the same benchmark.
The release also updates the language models used in testing. MLPerf Client v2.0 upgrades Phi 3.5 mini instruct to Phi 4 Mini Instruct and introduces Qwen 3 8B as an experimental test.
Another task in the benchmark is Intermediate Summarization, which uses an input prompt of roughly 4K tokens. With responsiveness and throughput reported together, the benchmark measures both how an AI system responds and how much work it can process.
A Bigger Test for AI Development Tools
AMD’s results and MLPerf Client v2.0 arrive with a shared focus: AI performance must be measured through real software tasks, not only code generation claims. AMD’s RSX experience shows that AI agents can begin with limited results and reach a 75 percent issue-resolution rate after development work.
MLPerf Client v2.0 gives agentic AI a broader testing structure, including Software Engineering and Data Analyst Agent scenarios. Its expanded categories and model updates create more ways to evaluate the systems that support these workflows.
The next stage of AI software development will center on how much of the engineering process agents can handle, from generated code to debugging and issue fixes. AMD’s 30 percent productivity boost, more than 80 percent AI-generated code in some components, and 75 percent RSX issue resolution show how quickly that process is moving.
Based on
- Most coding agent benchmarks skip large-scale refactoring. Not this one. — thenewstack.io
- AME Agent Swarms Quietly Rewrite the Workflow – IEEE Spectrum — spectrum.ieee.org
- MLPerf Client v2.0 Expands AI PC Benchmarking with Image Generation and Agentic AI | Markets Insider — markets.businessinsider.com
