The article presents a function making use of global variables, declares it bad, and then proposes that the solution is functional programming.
There's nothing wrong with promoting functional programming, but the implication that all non-FP code is hard to test and/or uses global state is naive.
It really isn't. Having worked for several decades on "both sides", this really is my experience. The functional side is better typed and has fewer side effects of this kind. It is more normal, as in more common, to have code work correctly as soon as it compiles. This is my lived experience having worked with Java, Scala, F# and Rust since 1999.
The author doesn't promote functional programming in the strictest sense, he just suggests to use functions that do not have side effects. He argues against changing state.
You can even do it in .NET using dependency injection (which is a kind of an OOP) concept: just supply dependencies as constructor parameters and do not change state in the service/handler.
You can follow the same concepts with ease using Go and no one will argue Go is a functional programming language.
So, the idea is not to do straight functional programming but to apply some lessons from the functional programming.
And I find the article to be very logic and even if I did not apply the same rules as the author, I empirically observed that coding agents shine where they don't have to argue much about state, the more state there is to keep track of, change and reason about, the harder it is for both humans and coding agents to do changes without breaking things.
And since I like the author's ideas I am willing to spend some tokens and put them to the test.
The only things I would add on top of this is modularization, either by using Vertical Slice Architecture or by designing a modular monolith where each module is isolated and has its own contract. Because, the less source code an agent has to reason about, the better are the results.
So my next greenfield experimental project will feature: compound engineering, TDD, SUPER and SPIRALS (as in the article) and modularization.
I really like the “functional core, imperative shell” approach, I try to use it whenever I can. I wish more non-FP languages had a way to mark a function as pure (and have this statically enforced).
My impression too. Most of what is in this post I've discovered almost entirely through clojure and I'm not even a clojure dev, just try to explore it in my free time. I did cmd + f in browser and searched for clojure pretty soon in to reading this.
I've explored Clojure after talking to Metabase about how it had benefited them. That said, it was years ago so I can't claim it influenced this work.
The framework was designed to be a language agnostic way of sharing best practices to bias agent behavior towards a more scalable end. I initially used it when I was working with a team to do some massive refactoring/clean up across the codebase. We didn't come to an acronym but similar principles and it was "testable" and easy to push back on PRs that weren't aligned with the principles.
That said, it may be interesting to see if I could replace all that context and just say -- "code it like you would with Clojure"
Functional programming also helped get ride of bugs before, and still people used other paradigms. Why would we change now? How to know that functional programming is indeed better for vibe coding?
It should be better for the reasons explained in the article. Pure functions require no context to understand. If they are typed, it's even simpler. LLMs perform badly on code that has lots of state and complex semantics. Those are hard to track.
In fact, synthesis of pure Haskell powered by SAT/SMT (e.g. Hoogle, Djinn, and MagicHaskeller) was already of some utility prior to the advent of LLMs. Furthermore, pure functions are also easy to test given that type signatures can be used for property-based test generation.
I think once all these components (LLMs, SAT/SMT, and lightweight formal methods) get combined, some interesting ways to build new software with a human-in-the-loop might emerge, yielding higher quality artifacts and/or enhancing productivity.
That's a very fair point. There are some publications showing lower performance for languages with less training data. I imagine it also applies to different paradigms. Most training code will be imperative and of lower quality.
it’s not discussed in this post but in another right after I discuss the modeling I was doing on tech debt and finding the game to improve agent outcomes was reducing context.
functional programming accomplishes that. I can’t claim it’s the only way, but it’s one that’s well understood in the community
What makes is special about "agentic development" vs reducing context requirements, reducing cognitive burden, etc for human development too? "A human developer builds a mental model of a codebase over months"—yeah, that makes onboarding to a codebase very time consuming, expensive, and error-prone.
So why is "better for agents" distinct from "better for humans"?
Agents can simply be told to write code in a functional style. They won’t complain. Think of it like a constraint system or proofs system. The agent can better reason about the code and side effects. Etc.
Agents are very good at following and validating constraints and hill climbing. This makes sense to me. Humans benefit too, but it is hard to get a bunch of humans to follow the style and maintain it over time.
Agents are useful because they don't inherit context from their parent context. They're basically "compaction" at a small scale. They succeed because context pollution create greater indeterminancy. The fact that you can spin up many of them is not primary benefit of them.
>finding the game to improve agent outcomes was reducing context
I think modularization will further reduce context. I am planning to play with your SUPER and SPIRALS idea and use modularization on top via Vertical Slice Architecture or modular monolith where each module is isolated and has a contract.
Because previously people had to write code themselves and they had personal preferences. Because the bugs were less when written by a human.
Now personal preferences doesn't matter as much since people won't change the code themselves too much. And because LLM can introduce lots of bugs and make a mess in the code unless you can restrict them somehow.
even if i would generally agree with the principles, no amount of markdown prompting is going to increase my confidence in agent's output and so i keep asking this question:
> what do you use for normative language to describe component boundary, function and cross-component interactions?
something i can feed into a deterministic system that will run a generative suite of tests (quickcheck/hypothesis/clojure.spec) that will either give me the confidence or give the agent the feedback.
OP / Author here, I started closer to where you are but ended up realizing I've led a few eng teams and was never satisfied with code quality. What I COULD be satisfied by was moving our metrics in the right direction. Testing coverage, use cases covered by E2E / integration tests, P99/backend efficiency metrics, cost of infrastructure and obviously user growth along with positive feedback from Users.
That said, I don't "vibe" because it creates great code I love reading, but I can monitor and move the same metrics I would if I was managing a team.
I also use code tours a bit, and one of my first tools I needed and built (intraview.ai) was to support this need to get deep in the code the Agents were claiming was ready to ship.
I've been having good luck with fairly autonomous LLM coding with the following rules:
* TypeScript everywhere with extreme enforcement of the type system.
* No "as" casts, no "any" declarations, all code must understand the shape of its data
* All boundaries validated using a typed validation library. Many use zod, I prefer tjs. I also have strictly typed pg and express wrappers.
* No files longer than 300 lines
* All of these rules are enforced by an eslint configuration that runs in a pre commit hook.
Global state and classes could also be removed via eslint rules, that would be interesting, though I haven't found it to be an issue in practice once the types are strictly enforced.
Can't fully enforce that one, but I require only my specific HTTP router / SQL library to be used, and disallow the import of express / pg libraries via eslint. I also disable the use of JSON.parse via eslint. Combined with strong language in the CLAUDE.md, I haven't seen it get around my guard rails.
So glad to see someone working on making fucking coding better — making LLMs write good code, not just vibe and wow, and not just pretend the code is magic.
The article presents a function making use of global variables, declares it bad, and then proposes that the solution is functional programming.
There's nothing wrong with promoting functional programming, but the implication that all non-FP code is hard to test and/or uses global state is naive.
It really isn't. Having worked for several decades on "both sides", this really is my experience. The functional side is better typed and has fewer side effects of this kind. It is more normal, as in more common, to have code work correctly as soon as it compiles. This is my lived experience having worked with Java, Scala, F# and Rust since 1999.
The author doesn't promote functional programming in the strictest sense, he just suggests to use functions that do not have side effects. He argues against changing state.
You can even do it in .NET using dependency injection (which is a kind of an OOP) concept: just supply dependencies as constructor parameters and do not change state in the service/handler.
You can follow the same concepts with ease using Go and no one will argue Go is a functional programming language.
So, the idea is not to do straight functional programming but to apply some lessons from the functional programming.
And I find the article to be very logic and even if I did not apply the same rules as the author, I empirically observed that coding agents shine where they don't have to argue much about state, the more state there is to keep track of, change and reason about, the harder it is for both humans and coding agents to do changes without breaking things.
And since I like the author's ideas I am willing to spend some tokens and put them to the test.
The only things I would add on top of this is modularization, either by using Vertical Slice Architecture or by designing a modular monolith where each module is isolated and has its own contract. Because, the less source code an agent has to reason about, the better are the results.
So my next greenfield experimental project will feature: compound engineering, TDD, SUPER and SPIRALS (as in the article) and modularization.
I really like the “functional core, imperative shell” approach, I try to use it whenever I can. I wish more non-FP languages had a way to mark a function as pure (and have this statically enforced).
Agreed. "Prefer pure functions, push side effects to the boundary (functional core, imperative)" is at the top of my CLAUDE.md these days.
This article is just describing Clojure. The SUPER principals are describing the native natural way of writing Clojure, no category theory needed.
My impression too. Most of what is in this post I've discovered almost entirely through clojure and I'm not even a clojure dev, just try to explore it in my free time. I did cmd + f in browser and searched for clojure pretty soon in to reading this.
OP here: thanks for chiming in.
I've explored Clojure after talking to Metabase about how it had benefited them. That said, it was years ago so I can't claim it influenced this work.
The framework was designed to be a language agnostic way of sharing best practices to bias agent behavior towards a more scalable end. I initially used it when I was working with a team to do some massive refactoring/clean up across the codebase. We didn't come to an acronym but similar principles and it was "testable" and easy to push back on PRs that weren't aligned with the principles.
That said, it may be interesting to see if I could replace all that context and just say -- "code it like you would with Clojure"
Have you tried that?
so does insisting on TDD.
I've known these things from the beginning.
Any extra restriction that still produces functional code ends up being great for LLMs to curb them deterministically.
Functional programming also helped get ride of bugs before, and still people used other paradigms. Why would we change now? How to know that functional programming is indeed better for vibe coding?
It should be better for the reasons explained in the article. Pure functions require no context to understand. If they are typed, it's even simpler. LLMs perform badly on code that has lots of state and complex semantics. Those are hard to track.
In fact, synthesis of pure Haskell powered by SAT/SMT (e.g. Hoogle, Djinn, and MagicHaskeller) was already of some utility prior to the advent of LLMs. Furthermore, pure functions are also easy to test given that type signatures can be used for property-based test generation.
I think once all these components (LLMs, SAT/SMT, and lightweight formal methods) get combined, some interesting ways to build new software with a human-in-the-loop might emerge, yielding higher quality artifacts and/or enhancing productivity.
Wouldnt a fair counter argument be, that llms have been trained on way less fu ctional code though?
Like they are trained on a LOT of js code -> good at js Way less functional code -> worse performance?
You can write functional-style code in many languages, as I have in JS and occasionally Python to great benefit.
For sure. I write functional style code in C# but it is not the same thing as writing OCaml or F#.
That's a very fair point. There are some publications showing lower performance for languages with less training data. I imagine it also applies to different paradigms. Most training code will be imperative and of lower quality.
OP Here:
it’s not discussed in this post but in another right after I discuss the modeling I was doing on tech debt and finding the game to improve agent outcomes was reducing context.
functional programming accomplishes that. I can’t claim it’s the only way, but it’s one that’s well understood in the community
What makes is special about "agentic development" vs reducing context requirements, reducing cognitive burden, etc for human development too? "A human developer builds a mental model of a codebase over months"—yeah, that makes onboarding to a codebase very time consuming, expensive, and error-prone.
So why is "better for agents" distinct from "better for humans"?
Agents can simply be told to write code in a functional style. They won’t complain. Think of it like a constraint system or proofs system. The agent can better reason about the code and side effects. Etc. Agents are very good at following and validating constraints and hill climbing. This makes sense to me. Humans benefit too, but it is hard to get a bunch of humans to follow the style and maintain it over time.
Agents are useful because they don't inherit context from their parent context. They're basically "compaction" at a small scale. They succeed because context pollution create greater indeterminancy. The fact that you can spin up many of them is not primary benefit of them.
Hmm? How does this relate to functional programming helping agents code better?
> The agent can better reason about the code and side effects. Etc. Agents are very good at following and validating constraints and hill climbing
You're treating an agent like its something other than an LLM with limited context. I'm just trying to surface a incorrect assumption.
You're just talking about LLMs. Whether they're a child of a primary LLM is irrelevent to your conception.
It'd be interesting if you posited that LLMs doing functional style could handle more context without service degradation.
>finding the game to improve agent outcomes was reducing context
I think modularization will further reduce context. I am planning to play with your SUPER and SPIRALS idea and use modularization on top via Vertical Slice Architecture or modular monolith where each module is isolated and has a contract.
>Why would we change now?
Because previously people had to write code themselves and they had personal preferences. Because the bugs were less when written by a human.
Now personal preferences doesn't matter as much since people won't change the code themselves too much. And because LLM can introduce lots of bugs and make a mess in the code unless you can restrict them somehow.
even if i would generally agree with the principles, no amount of markdown prompting is going to increase my confidence in agent's output and so i keep asking this question:
> what do you use for normative language to describe component boundary, function and cross-component interactions?
something i can feed into a deterministic system that will run a generative suite of tests (quickcheck/hypothesis/clojure.spec) that will either give me the confidence or give the agent the feedback.
OP / Author here, I started closer to where you are but ended up realizing I've led a few eng teams and was never satisfied with code quality. What I COULD be satisfied by was moving our metrics in the right direction. Testing coverage, use cases covered by E2E / integration tests, P99/backend efficiency metrics, cost of infrastructure and obviously user growth along with positive feedback from Users.
That said, I don't "vibe" because it creates great code I love reading, but I can monitor and move the same metrics I would if I was managing a team.
I also use code tours a bit, and one of my first tools I needed and built (intraview.ai) was to support this need to get deep in the code the Agents were claiming was ready to ship.
I've been having good luck with fairly autonomous LLM coding with the following rules:
Global state and classes could also be removed via eslint rules, that would be interesting, though I haven't found it to be an issue in practice once the types are strictly enforced.How do you enforce the use of validation library with eslint?
Can't fully enforce that one, but I require only my specific HTTP router / SQL library to be used, and disallow the import of express / pg libraries via eslint. I also disable the use of JSON.parse via eslint. Combined with strong language in the CLAUDE.md, I haven't seen it get around my guard rails.
So glad to see someone working on making fucking coding better — making LLMs write good code, not just vibe and wow, and not just pretend the code is magic.
close
[dead]