We present a determination of the optical/UV AGN luminosity function and its evolution, based on a large sample of faint (R < 24) QSOs identified in the COMBO-17 survey. Using multi-band photometry in 17 filters within 350 nm &LSIM; λ(obs) &LSIM; 930 nm, we could simultaneously determine photometric redshifts with an accuracy of σ(z) < 0.03 and obtain spectral energy distributions. The redshift range covered by the sample is 1.2 < z < 4.8, which implies that even at z similar or equal to 3, the sample reaches below luminosities corresponding to M-B = -23, conventionally employed to distinguish between Seyfert galaxies and quasars. We clearly detect a broad plateau-like maximum of quasar activity around z similar or equal to 2 and map out the smooth turnover between z similar or equal to 1 and z similar or equal to 4. The shape of the LF is characterised by some mild curvature, but no sharp "break" is present within the range of luminosities covered. Using only the COMBO-17 data, the evolving LF can be adequately described by either a pure density evolution (PDE) or a pure luminosity evolution (PLE) model. However, the absence of a strong L*-like feature in the shape of the LF inhibits a robust distinction between these modes. We present a robust estimate for the integrated UV luminosity generation by AGN as a function of redshift. We find that the LF continues to rise even at the lowest luminosities probed by our survey, but that the slope is sufficiently shallow that the contribution of low-luminosity AGN to the UV luminosity density is negligible. Although our sample reaches much fainter flux levels than previous data sets, our results on space densities and LF slopes are completely consistent with extrapolations from recent major surveys such as SDSS and 2QZ.