An AI agent that shrinks neural networks without wrecking their fairness just got a proof of concept.
Researchers built FairCompressAgent (FCA), a system that combines three model-compression tricks - pruning, quantization, and low-rank factorization - behind a single interface. A language-model planner picks which combination to try based on a model's profile and past results, then an execution layer runs the compression, fine-tunes the model, tests it, and checks whether it meets accuracy and fairness targets. Users can change those targets mid-process, and the system reports exactly which requirement it fails to hit if it can't satisfy the request. In tests on a skin-condition image dataset (Fitzpatrick-17k) using a VGG-11 model, FCA cut inference storage by 59.54 percent, raised average precision from 0.5141 to 0.5233, and reduced a fairness metric called equalized opportunity from 0.2251 to 0.2168 - while needing fewer test runs (7.33 versus 12) than a one-shot alternative.
The real story here isn't the numbers, it's the workflow. Model compression for edge devices like FPGAs usually treats fairness as an afterthought you check once at the end, if at all. FCA builds the fairness check into the search itself, and lets you renegotiate the tradeoff live instead of starting over.
That said, this is one dataset, one architecture, and one paper's own benchmarks. Calling a constrained search-and-select loop "agentic" is very 2026, but the underlying idea - make the tradeoffs explicit and interactive - is worth stealing regardless of the label.