Lisps are "wizard" languages: the runtime semantic is kept close to the syntax, which also means that you can rapidly extend the syntax to solve a problem. This quality is true of Forth as well, and shares some energy with APL and its "one symbol for one behavior" flavor. With respect to their metaprogramming, the syntactical approach hands you a great foot-gun in that you can design syntax that is very confusing and specific to your project, which nobody else will be able to ramp up on.

But Algols, including Python, are "bureaucrat" languages: rather than condensing the syntax to be the exact specification of the program, they define a right way to form the expression, and then the little man inside the compiler rubber stamps it when you press the run button. In other words, they favor defining a semantics and then adding a syntax around that, which means that they need more syntax, it's harder to explain the behavior of a syntactical construction, and it's harder to extend to have new syntax. But by being consistent in a certain form-filling way, they enable a team to collaborate and hand off relatively more code.

IMHO, a perfectly reasonable approach I'm exploring now for personal work is to have a Lisp(or something that comes close enough in size, dynamic behavior, and convenience, like Lua) targeting a Forth. The Forth is there to be a stack machine with a REPL. You can extend the Forth upwards a little bit because it can metaprogram, or downwards to access the machine. It is better for development than a generic bytecode VM because it ships in a bootstrappable form, with everything you need for debugging and extension - it is there to be the layer that talks to the machine, as directly as possible, so nothing is hidden behind a specialized protocol. And you can use the Lisp to be the compiler, to add the rubber-stamping semantics, work through resource tracking issues, do garbage collection and complicated string parsing, and generate "dumb" Forth code where it's called for. That creates a nice mixture of legibility and configurability, where you can address the software in a nuts-and-bolts way or with "I want an algorithm generating this functionality".