We present new near-infrared images of z > 0.8 radio galaxies from the flux
-limited 7C-III sample of radio sources for which we have recently obtained
almost complete spectroscopic redshifts. The 7C objects have radio luminos
ities approximate to 20 times fainter than 3C radio galaxies at a given red
shift. The absolute magnitudes of the underlying host galaxies and their sc
ale sizes are only weakly dependent on radio luminosity. Radio galaxy hosts
at z similar to 2 are significantly brighter than the hosts of radio-quiet
quasars at similar redshifts and the recent model AGN hosts of Kauffmann &
Haehnelt. There is no evidence for strong evolution in scale size, which s
hows a large scatter at all redshifts. The hosts brighten significantly wit
h redshift, consistent with the passive evolution of a stellar population t
hat formed at z greater than or similar to 3. This scenario is consistent w
ith studies of host galaxy morphology and submillimeter continuum emission,
both of which show strong evolution at z greater than or similar to 2.5. T
he lack of a strong "redshift cutoff" in the radio luminosity function to z
> 4 suggests that the formation epoch of the radio galaxy host population
lasts greater than or similar to 1 Gyr, from z greater than or similar to 5
to z similar to 3. We suggest these facts are best explained by models in
which the most massive galaxies and their associated AGN form early because
of high baryon densities in the centers of their dark matter haloes.