Does a non-lagrangian field theory have a stress-energy tensor?

Most theories do not have a conserved energy-momentum tensor, regardless of whether they are Lagrangian or not. For example you need locality and Lorentz invariance. When you have those, you can define the energy-momentum tensor via the partition function (which always exists, it basically defines the QFT): $$ langle T_{munu}rangle:=frac{delta}{delta g^{munu}}Z[g] $$ You can define arbitrary correlation functions of $T$ by including insertions. And these functions define the operator $T$ itself. So $T$ is defined whenever $Z$ is a (differentiable) function of the metric, i.e., when you have a prescription to probe the dependence of the theory on the background metric. And this prescription is part of the definition of the theory: in order to define a QFT you must specify how the partition function is to be computed, for an arbitrary background. If you cannot or do not want to specify $Z$ for arbitrary $g$, then the derivative cannot be evaluated and $T$ is undefined. (This is not an unreasonable situation, e.g. I may be dealing with a theory that is anomalous and is only defined for a special set of metrics, e.g. Kähler. This theory does not make sense for arbitrary $g$, so I may not be able to evaluate $Z[g]$ for arbitrary $g$, and therefore $T$ may not exist).

If the theory admits an action, then the dependence of $Z$ on $g$ is straightforward: it is given by whatever the path-integral computes. If the theory does not admit an action, then you must give other prescriptions by which to compute $Z$. This prescription may or may not include a definition for arbitrary $g$; if it does not, then $T$ is in principle undefined.

But anyway, for fun consider the following very explicit example: $mathcal N=3$ supersymmetry in $d=4$. This theory is known to be non-lagrangian. Indeed, if you write down the most general Lagrangian consistent with $mathcal N=3$ SUSY, you can actually prove it preserves $mathcal N=4$ SUSY as well. So any putative theory with strictly $mathcal N=3$ symmetry cannot admit a Lagrangian. Such a theory was first constructed in arXiv:1512.06434, obtained almost simultaneously with arXiv:1512.03524. This latter paper analyses the consequences of the anomalous Ward identities for all symmetry currents, in particular the supercurrent and the energy-momentum tensor.