Researchers have built a system where AI agents write and rewrite the code for optimization algorithms, no human algorithm designer required.
The approach targets Meta-Black-Box Optimization, or MetaBBO, a method that uses a learnable policy to guide optimization tasks like scheduling or parameter tuning. Until now, MetaBBO systems were handcrafted for each problem type, which limited how good they could get. The new framework uses two agents instead: a task agent that keeps rewriting the optimizer's codebase, and a hyper agent that modifies the task agent and itself based on how each version performs. The result, evaluated against current human-designed MetaBBO baselines, is optimizer code that the researchers say outperforms those baselines and adapts quickly to new problem domains.
This matters because it is a concrete example of AI systems improving the tools used to build other AI systems, not just generating app code or chat responses. If self-improving design loops like this generalize beyond optimization algorithms, it raises real questions about how much of algorithm design work still needs a human in the loop.
The claims come from the paper's own benchmarks against existing baselines, so independent replication will matter before anyone rewrites their optimization pipeline around it.