I loved Python for a long time.
The abstractions. Custom DSLs with that gamified syntax feel. Metaclasses, descriptors, decorators that bend the runtime. All the magic you could pull off in a hundred lines and feel clever doing it. That was the fun.
Not anymore.
The agentic coding shift
When I switched to agentic coding, something changed. About 95% of the code an LLM produces in Python defaults to the same shape: FastAPI + Pydantic AI slop with no soul. Map a Pydantic model to another Pydantic model. Validate the validated. Wrap a wrapped thing. Type the dynamic. Build illusionary safety on top of a duck-typed runtime and call it done.
It is not fun to review. It is not fun to extend. It is just... there.
I went back to Go
I switched my default to Go. Not because of the type system. Not for performance. Not for compile times. Not for any of the usual talking points.
I switched because of variance. Or, if you want to be a snob about it, perplexity of the output.
I used Go a lot in the past. Had a proper love-hate relationship with it. But starting a new project today, my default is no longer Python. LLMs produce decent, high-quality Go. There are fewer ways to be clever. Fewer hidden traps. Fewer "let me invent a framework" detours.
Some haters will tell you Rust or Zig is the right move. I would then counter with Haskell, and we would argue forever while nothing ships. Go just gets the job done. Easy to review, somehow easier to test, and feels 10x faster and more stable in practice.
The work moved up a level
When the model produces reliable output, the work shifts. It is no longer about syntax that compiles. It is about meta-position and plan.
Where does this service sit. What is the contract. What is the failure mode. What is the smallest thing that proves the assumption. The boring questions that matter.
The code becomes the cheapest part.
A quiet revision
I used to think Go was a language for mediocre developers who struggled with deeper programming concepts. Generics-allergic. Pattern-matching-less. No sum types. No expression-everything.
I was wrong about who it is for.
Go turns out to be the perfect target for a mediocre LLM code printer that just implements your intent into code. The constraints that felt like limitations to a human are exactly the constraints that keep an LLM honest. Fewer degrees of freedom equals fewer ways to spiral.
You can sprinkle in property-based testing if you want to feel cool about it.
What it looks like in practice
- New project?
go mod initbeforeuv init. - Need a CLI, a service, a worker, a glue script? Go.
- Need a notebook, a one-off data thing, an ML pipeline? Sure, still Python.
- Need to be impressed by yourself? Open a Haskell repo on a Sunday.
The magic was fun. The boring is faster.