IEEE Spectrum breed · Technologie

AMD's AI Leap: From Code Generation to Agent Swarms in Software Development

AMD has achieved a 30% productivity boost in software development through AI, surpassing its initial target. The company is now looking beyond current AI capabilities, envisioning collaborative AI agent swarms that can independently discover solutions, fundamentally reshaping the software development lifecycle (SDLC).

AMD initially aimed for a 25% productivity increase from AI in software development over two to three years, focusing on tasks like code generation, problem triaging, debugging, and testing. However, rapid advancements in large language models (LLMs) have accelerated this progress, leading to a current 30% productivity gain and prompting a re-evaluation of the SDLC itself.

The company has integrated AI into every stage of the SDLC. AI agents now analyze problem reports, group similar issues, identify code requiring modification, debug code, generate unit tests, and prepare documentation for review and release. This has led to AI generating over 20% of the codebase, with some components exceeding 80% AI-generated code.

Current AI agents are trained to mimic human workflows, which limits their potential. AMD anticipates the next major transformation will involve collaborative AI agent swarms that can independently identify and develop solutions. Instead of detailed instructions, engineers will define the problem, desired outcomes, and constraints, allowing agents to determine the optimal path.

These agent swarms will generate, evaluate, and refine multiple solutions in parallel, validating correctness, measuring performance, and testing trade-offs. The goal is for these agents to rewrite the SDLC rather than just enhance existing steps. To achieve this, agents need to learn collaboratively from each other, reusing successful strategies across projects.

AMD is already experimenting with multi-agent workflows and developing internal systems for the next generation of AI-driven software engineering. An example is their AI-driven effort to resolve issues in Radeon Software eXperience (RSX), where AI agents have increased issue resolution rates from 6% to over 75% by refining objectives and iteratively exploring solutions.

The company emphasizes that AI is intended to increase productivity and quality, enabling employees to focus on higher-value work, not to reduce headcount. AMD is investing in AI education and training to ensure its workforce can confidently leverage AI, with engineers shifting their focus towards defining specifications, validating outcomes, and making strategic decisions.

AI-samenvatting op basis van de bron.