This paper deals with robust consensus of uncertain linear multi-agent systems via dynamic output-feedback protocols. Agents are assumed to have identical nominal linear time-invariant dynamics but are subject to heterogeneous additive stable unmodelled dynamics. We consider a dynamic output-feedback protocol of the observer-based type without direct exchange of local controller states. Sufficient conditions ensuring robust consensus for all uncertainties with a given H_∞ norm bound are obtained from robust control theory, based on which a design algorithm is developed for the protocol gains to achieve closed-loop robust consensus. The proposed method can simultaneously cope with heterogeneous uncertainties and directed communication graphs. The advantages of the proposed design method are finally illustrated by a numerical example.