AI Now Rewrites Its Own Problem-Solving Recipes Mid-Task

When computers hunt for the best answer among millions of options — the cheapest delivery route, the strongest new material, the fastest factory schedule — they follow step-by-step instructions called an algorithm. Today those instructions are usually written by humans, or by AI in a one-off session before the job starts. The problem: as the task shifts midway, a fixed recipe can't adapt. That's like navigating a road trip using directions written before anyone knew about the traffic jam.

A research team has now built a system that treats the instructions themselves as something to revise on the fly. Their framework, OnDesign, runs several AI agents — software that can reason and write code — and each proposes a slightly different approach. The agents reconcile their competing ideas, produce a working piece of code, run it, then use what happened to shape the next round. No separate training stage is needed; the AI learns what works by watching what it just did.

In tests across six standard problem sets — including Bayesian optimization (a smart way to search when each test is expensive) and evolutionary optimization (mimicking survival of the fittest to improve solutions) — OnDesign outperformed both traditional optimizers and earlier AI design methods. The team also applied it to a real engineering problem, and the gains held up as problems grew larger, which is where conventional methods usually struggle.

The catch: this is early-stage research, not something you can use today. It requires serious computing power, the AI-written algorithms can be hard for humans to inspect or trust, and the tests were mathematical and engineering puzzles rather than messy real-world situations. Still, the direction matters. If AI can keep improving its own methods mid-task, the cost of solving the optimization problems behind shipping logistics, drug discovery, and chip design could fall.