Abstract Bacterial gene transfer agents (GTAs) are small virus-like particles that package DNA fragments and inject them into cells. They are encoded by gene clusters resembling defective prophages, with genes for capsid head and tail components. These gene clusters are usually assumed to be maintained by selection for the benefits of GTA-mediated recombination, but this has never been tested. We rigorously examined the potential benefits of GTA-mediated recombination, considering separately transmission of GTA-encoding genes and recombination of all chromosomal genes. In principle GTA genes could be directly maintained if GTA particles spread them to GTA − cells often enough to compensate for the loss of GTA-producing cells. However careful bookkeeping showed that losses inevitably exceed gains for two reasons. First, cells must lyse to release particles to the environment. Second, GTA genes are not preferentially replicated before DNA is packaged. A simulation model was then used to search for conditions where recombination of chromosomal genes makes GTA + populations fitter than GTA − populations. Although the model showed that both synergistic epistasis and some modes of regulation could generate fitness benefits large enough to overcome the cost of lysis, these benefits neither allowed GTA + cells to invade GTA − populations, nor allowed GTA + populations to resist invasion by GTA − cells. Importantly, the benefits depended on highly improbable assumptions about the efficiencies of GTA production and recombination. Thus, the selective benefits that maintain GTA gene clusters over many millions of years must arise from consequences other than transfer of GTA genes or recombination of chromosomal genes.