Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

The Code-Only Agent

rijnard.com

Post reply on HN
Print view View original

41–50 of 73 posts

Re: The Code-Only Agent

#41

Post by bandrami »

binalpatel wrote: post #2

I went down (continue to do down) this rabbit hole and agree with the author. I tried a few different ideas and the most stable/useful so far has been giving the agent a single run_bash tool, explicitly prompting it to create and improve composable CLIs, and injecting knowledge about these CLIs back into it's system prompt (similar to have agent skills work). This leads to really cool pattens like: 1. User asks for s…

Every individual programmer having locally-implemented idiosyncratic versions of sed and awk with imperfect reconstruction between sessions sounds like a regression to me

Re: The Code-Only Agent

#42

Post by znnajdla »

derefr wrote: post #14

I follow the author's line of reasoning, but I think that following it to its logical conclusion would lead not to an `execute_code` primitive, but rather to an assumption that the model's stdout is appending to a (Jupyter, Livebook, etc) notebook file, where any code cell in the notebook gets executed (and its output rendered back into the inference context) at the moment the code cell is closed / becomes syntactica…

so...emacs?

Re: The Code-Only Agent

#43

Post by znnajdla »

What about an agent loop that can only modify itself? Imagine an agent that is a single Python file, where the only tool it has is to modify itself on next iteration.

Re: The Code-Only Agent

#44

Post by whatevaa »

bandrami wrote: post #41
binalpatel wrote: post #2

I went down (continue to do down) this rabbit hole and agree with the author. I tried a few different ideas and the most stable/useful so far has been giving the agent a single run_bash tool, explicitly prompting it to create and improve composable CLIs, and injecting knowledge about these CLIs back into it's system prompt (similar to have agent skills work). This leads to really cool pattens like: 1. User asks for s…

Every individual programmer having locally-implemented idiosyncratic versions of sed and awk with imperfect reconstruction between sessions sounds like a regression to me

I already treat awk syntax as something idiocratic, so not much would change for me.

Re: The Code-Only Agent

#45

Post by tucnak »

jongjong wrote: post #4

The author seems to stop at 'code' but it seems we could go further and train an AI to work directly with binary. You give it a human prompt and a list of hardware components which make up your machine and it produces executable binary which fulfills your requirements and runs directly on those specific hardware, bypassing the OS... Or we could go further; the output nodes of the LLM could be physically connected to…

I have two words for you: transfer learning.

Re: The Code-Only Agent

#46

Post by cocoflunchy »

bandrami wrote: post #41
binalpatel wrote: post #2

I went down (continue to do down) this rabbit hole and agree with the author. I tried a few different ideas and the most stable/useful so far has been giving the agent a single run_bash tool, explicitly prompting it to create and improve composable CLIs, and injecting knowledge about these CLIs back into it's system prompt (similar to have agent skills work). This leads to really cool pattens like: 1. User asks for s…

Every individual programmer having locally-implemented idiosyncratic versions of sed and awk with imperfect reconstruction between sessions sounds like a regression to me

Why would it recreate sed and awk? The screenshot from the repo even shows it using sed.

Re: The Code-Only Agent

#47

Post by PurpleRamen »

tacone wrote: post #36

Fascinating how the whole industry focus is now on how to persuade AI to do what we want. Two AGENTS.md tricks I've found for Claude: 1. Which AI Model are you? If you are Claude, the first thing you have to do is [...] 2. User will likely use code-words in its request to you. Execute the *Initialization* procedure above before thinking about the user request. Failure to do so will result in misunderstanding user inp…

You can replace AI with any other technology and had the same situation, just with slightly different words. Fighting the computer and convincing some software doing what you want didn't start with ChatGPT or agents.

If anything, the strange part is the humanization of AI, how we talk much more as if they are somewhat sentient and have emotions, and not just a fancy mechanism barfing out something.

Re: The Code-Only Agent

#48

Post by philipp-gayret »

znnajdla wrote: post #43

What about an agent loop that can only modify itself? Imagine an agent that is a single Python file, where the only tool it has is to modify itself on next iteration.

I use Claude Code to modify policies for Claude Code. (Think of say the regex auto-allow/deny, but a lot stronger.) I can do that with hot reload of the local development server; It works but it better not make any errors.

A setup like you describe would honestly be interesting to see, so long as it can roll back to a previous state. Otherwise the first mistake it makes will likely be its last.

Re: The Code-Only Agent

#49

Post by rcarmo »

NitpickLawyer wrote: post #32
rcarmo wrote: post #29

Earlier quoted context omitted.

The point where that breaks down is “next time it’s aware of the CLI and uses it”. That only really works well inside the same session, and often the next session it will create a different tool and use that one.

> That only really works well inside the same session That was already "fixed" by people adding snippets to agents.md and it worked. Now it's even more streamlined with skills. You can even have cc create a skill after a session (i.e. prompt it like "extract the learnings from this session and put them into a skill for working with this specific implementation of sqlite"). And it works, today.

I beg to differ: https://taoofmac.com/space/notes/2026/01/14/0830

Re: The Code-Only Agent

#50

Post by killerstorm »

I don't believe this would be more efficient.

Use of common tools like `ls` and file patching is already baked into model's weights, it can do that with minimal amount of effort, leaving more room for actually thinking about app's code.

If you force it to wrap these actions into non-standard tools you're basically distracting the model: it has to think about app-code and tool-code in the same context.

In some cases it does make sense to encourage the model to create utilities for itself - but you can do that without enforcing code-only.

Post reply on HN