What people should not do is cascade ciphers, or, even worse, build systems that allow for cipher negotiation. Every system that negotiates has been a disaster.
20+ years of issues with SSL and SSH, both in the standards and implementations, has taught us that the costs of supporting algorithm agility is much more than the gains. Borderline intolerable, in fact. Moreover, our two-decade track record of cryptanalysis and predictions of future breakthroughs has been very good.
We're much better able to predict the costs+benefits of a particular cryptographic primitive than we are of the protocols and software we build. If a cryptographic primitive is assessed as good enough, it's no longer considered desirable or even reasonable to hedge that assessment and permit substitution of the primitive during the lifetime of the protocol. The resultant complexity is much more likely to be the source of future problems than the primitive.
Ex: when TLS clients and servers commonly supported 3DES, RC4, and AES, it wasn't universally painful to disable 3DES and RC4 (although if you support(ed) embedded clients without AES support, it's still painful). I think ChaCha looks to be a second good option that's gaining support.