Not yet reviewed by Pith; the record is open.

Spitzer IRAC selection is a powerful tool for identifying luminous AGN. For deep IRAC data, however, the AGN selection wedges currently in use are heavily contaminated by star-forming galaxies, especially at high redshift. Using the large samples of luminous AGN and high-redshift star-forming galaxies in COSMOS, we redefine the AGN selection criteria for use in deep IRAC surveys. The new IRAC criteria are designed to be both highly complete and reliable, and incorporate the best aspects of the current AGN selection wedges and of infrared power-law selection while excluding high redshift star-forming galaxies selected via the BzK, DRG, LBG, and SMG criteria. At QSO-luminosities of log L(2-10 keV) (ergs/s) > 44, the new IRAC criteria recover 75% of the hard X-ray and IRAC-detected XMM-COSMOS sample, yet only 38% of the IRAC AGN candidates have X-ray counterparts, a fraction that rises to 52% in regions with Chandra exposures of 50-160 ks. X-ray stacking of the individually X-ray non-detected AGN candidates leads to a hard X-ray signal indicative of heavily obscured to mildly Compton-thick obscuration (log N_H (cm^-2) = 23.5 +/- 0.4). While IRAC selection recovers a substantial fraction of luminous unobscured and obscured AGN, it is incomplete to low-luminosity and host-dominated AGN.