New AI Fixes Software Bugs by Assembling Its Own Repair Crew

Every app, website, and banking system you use runs on code, and that code breaks. When it does, programmers hunt for the mistake — a process called debugging. Increasingly, companies hand that job to AI. But today's AI debugging tools work like a repair crew with a fixed number of workers: it doesn't matter whether the problem is a wobbly doorknob or a collapsed roof, the same team shows up.

A group of researchers from Tunisia introduced ASAD, which throws out that one-size-fits-all rule. ASAD starts by reading the broken code, then decides for itself how many AI helpers to deploy, what each one should specialize in, and how they should work together. A coordinator AI runs the show, planning and checking its own work as it goes. A typo gets a single fast pass. A tangled security flaw gets a purpose-built team of specialists.

The payoff is measurable. Tested against three standard sets of real-world bugs and run on several AI models, including DeepSeek, Qwen, and GPT-5, ASAD fixed 12-20% more bugs than the common 'think step by step' approach, and beat fixed-team AI systems by 4-9% in accuracy. Just as important, it used about a third fewer AI helpers overall — which matters because every extra AI burns money, time, and electricity.

So what does this mean for you? Software failures already cost the global economy an estimated hundreds of billions of dollars a year, and every hour a payment system or hospital record system is down, real people are inconvenienced or put at risk. AI that fixes bugs quickly and cheaply means smaller companies — not just tech giants — can afford to keep their products reliable. It also hints at where AI is heading generally: instead of one giant model doing everything, we get small teams of AI that scale up only when the job demands it.