Multi-wavelength studies of deep radio fields show a composite population of star-forming galaxies, radio-quiet and radio-loud AGNs, with the formers dominating at the lowest flux densities (< 100 microJy). However, the exact mixture between these types of radio sources is still matter of debate. The most direct way to identify faint AGN-driven radio emission is the detection of embedded radio cores in the host galaxies, through ultra-deep and very high resolution radio observations. This would open the perspective of studying the whole AGN population in the radio band, including the radio-quiet component traditionally selected at other wavelengths (opt/IR/X-ray). Assessing the faint AGN component in deep radio fields, will provide an important tool to understand the role of nuclear activity in distant galaxies and its possible co-evolution with star-formation processes, as radio wavelengths are not affected by dust extinction and/or gas absorption. In my talk I will report about the e-MERLIN Galaxy Evolution Survey (eMERGE, PI: Muxlow), a legacy project which aims at undertaking a spatially-resolved study of AGN and star formation processes up to high redshift in a 30 arcmin diameter field in the GOODS-N region, through ultra-deep (sub-microJy rms), sub-arcsec (50-500 mas) imaging at 1.4 and 5 GHz, using combined JVLA and eMERLIN observations. I will focus on the 5 GHz JVLA mosaic observations and catalogue of GOODS-N (94 sources), in the framework of the eMERGE project, and on the study of a larger sample of GOODS-N galaxies (300 objects) selected at 1.4 GHz to constrain the presence of AGN cores in moderate-to-high redshift (1<z<5) galaxies, via radio spectra-morphological analysis with the additional help of multi-wavelength information.