Multi-agent systems on nonlinear spaces sometimes fail to synchronize. This is usually attributed to the initial configuration of the agents being too spread out, the graph topology having certain undesired symmetries, or both. Besides nonlinearity, the role played by the geometry and topology of the nonlinear space is often overlooked. This paper concerns two gradient descent flows of quadratic disagreement functions on general Riemannian manifolds. One system is intrinsic while the other is extrinsic. We derive necessary conditions for the agents to synchronize from almost all initial conditions when the graph used to model the network is connected. If a Riemannian manifold contains a closed curve of locally minimum length, then there is a connected graph and a dense set of initial conditions from which the intrinsic system fails to synchronize. The extrinsic system fails to synchronize if the manifold is multiply connected. The extrinsic system appears in the Kuramoto model on S 1 , rigid-body attitude synchronization on SO(3), the Lohe model of quantum synchronization on the n-sphere, and the Lohe model on U(n). Except for the Lohe model on the n-sphere where n ∈ N\{1}, there are dense sets of initial conditions on which these systems fail to synchronize. The reason for this difference is that the n-sphere is simply connected for all n ∈ N\{1} whereas the other manifolds are multiply connected.