LLM-generated code that seemingly went without much review or design is always such an interesting dive into just how bloated you can make code. In this repository, multiple files are close to 10K LOC, one file contains a switch statement that has so many case statements it spans more than 1000 lines, and lots of other fun stuff.
I guess it depends on the model you're trying to use, but seems most of them prefer smaller codebases, they work a lot better with less code, which kind of makes sense. With that in mind, I'd probably aim for something way smaller to bootstrap a self-improving agent. Then I'd use this "Prime Agent" as an example to my self-improving agent for what it should not evolve to.
For everyone downvoting: It's literally a story about the creation of mankind's first artificial general intelligence, Prime Intellect, and the consequences of that discovery.
I feel that more warning is needed for this book. Everything you’ve stated is true, but the book’s first chapter contains some of the most disturbing depictions of stuff I don’t know I can type here on HN. Additionally, the last chapter put a really bad taste in my mouth.
That said, the stuff that deals with AI and its implications on the universe was great and more relevant than ever. The way PI works is pretty similar to how we use subagents to tackle large repos!
I built one of these RLM harnesses and a local MCP server along with logging, memories, and project rules based on directories. It worked great for a while but the foundational models have largely caught up to the point where they don't need this harness anymore. At least for my use cases. I can basically just store context in .md in the directories we work out of together and accomplish what I need.
I went the skills route - skill improvement skill and a rule to use it always. Basically it's directed that if anything causes more than a hop of thinking - failure - try something else it should flag that it needs to learn it as a skill so it never does it other than first shot again or if an existing skill fails improve it after it solves whatever problem. it then syncs the files to a shared location and updates the version and also pulls new skills. this let's a team use it or you have multiple workstations.
The core idea of the RLM paper is to make a regular LLM act more like a coding agent - offload context to something external that needs to be explicitly queried instead of filling up valuable context. The "recursion" part of the paper really only wins because they use a top-tier model for the root agent, and cheaper models for the sub-agents.
Prime Agent just took the RLM idea (which is really just an academic view on how coding agents have always worked) and then added this "continual harness" idea. This part isn't super well described in the blog post, but includes some message passing between the agents, and the ability to share code.
Overall I chalk it up as neat, but not revolutionary. Another version of what most of these systems are already doing.
It'll be really interesting when they run RL training on the harness self-improvement loop. I've tried using LLMs for harness engineering, but it often creates too much bloat that weighs things down in the end. Guessing it's just not something the models are tuned to do by default.
Curious if anyone's tried using RL for harness engineering? I think we're still pretty far away from the optimal harness, especially when it comes to long-context memory management.
It’s a self-improving harness, and ARC-AGI-3 is explicitly a few-shot benchmark. It’s likely that it gave itself more than the maximum number of tries to learn the games, or even hardcoded the answers.
There’s a lot of improvement to be had from the benchmark harnesses, but sometimes, like with ARC-AGI-3, the limitations are intentional.
LLM-generated code that seemingly went without much review or design is always such an interesting dive into just how bloated you can make code. In this repository, multiple files are close to 10K LOC, one file contains a switch statement that has so many case statements it spans more than 1000 lines, and lots of other fun stuff.
I guess it depends on the model you're trying to use, but seems most of them prefer smaller codebases, they work a lot better with less code, which kind of makes sense. With that in mind, I'd probably aim for something way smaller to bootstrap a self-improving agent. Then I'd use this "Prime Agent" as an example to my self-improving agent for what it should not evolve to.
> one file contains a switch statement that has so many case statements it spans more than 1000 lines
Probably best to leave YandereDev's code out of the training data.
Bash is all you need.
https://minimal-agent.com/
https://localroger.com/prime-intellect/mopiidx.html
For everyone downvoting: It's literally a story about the creation of mankind's first artificial general intelligence, Prime Intellect, and the consequences of that discovery.
I feel that more warning is needed for this book. Everything you’ve stated is true, but the book’s first chapter contains some of the most disturbing depictions of stuff I don’t know I can type here on HN. Additionally, the last chapter put a really bad taste in my mouth.
That said, the stuff that deals with AI and its implications on the universe was great and more relevant than ever. The way PI works is pretty similar to how we use subagents to tackle large repos!
yea the beginning is pretty gory / horrific. it almost made me stop reading but you can't deny the world building is.. unique
I built one of these RLM harnesses and a local MCP server along with logging, memories, and project rules based on directories. It worked great for a while but the foundational models have largely caught up to the point where they don't need this harness anymore. At least for my use cases. I can basically just store context in .md in the directories we work out of together and accomplish what I need.
I went the skills route - skill improvement skill and a rule to use it always. Basically it's directed that if anything causes more than a hop of thinking - failure - try something else it should flag that it needs to learn it as a skill so it never does it other than first shot again or if an existing skill fails improve it after it solves whatever problem. it then syncs the files to a shared location and updates the version and also pulls new skills. this let's a team use it or you have multiple workstations.
The core idea of the RLM paper is to make a regular LLM act more like a coding agent - offload context to something external that needs to be explicitly queried instead of filling up valuable context. The "recursion" part of the paper really only wins because they use a top-tier model for the root agent, and cheaper models for the sub-agents.
Prime Agent just took the RLM idea (which is really just an academic view on how coding agents have always worked) and then added this "continual harness" idea. This part isn't super well described in the blog post, but includes some message passing between the agents, and the ability to share code.
Overall I chalk it up as neat, but not revolutionary. Another version of what most of these systems are already doing.
It'll be really interesting when they run RL training on the harness self-improvement loop. I've tried using LLMs for harness engineering, but it often creates too much bloat that weighs things down in the end. Guessing it's just not something the models are tuned to do by default.
Curious if anyone's tried using RL for harness engineering? I think we're still pretty far away from the optimal harness, especially when it comes to long-context memory management.
A write up on RLM (Recursive Language Models) by one of the authors of the RLM paper:
https://alexzhang13.github.io/blog/2025/rlm/
It is impressive that it (almost) saturates ARC-AGI-3, https://x.com/PrimeIntellect/status/2085087000764568010.
I am curious - how does it fare for other benchmarks, or everyday programming?
PrimeIntelect is not on official ARC-AGI-3 leaderboard: https://arcprize.org/leaderboard
Good to know!
Is it that it wasn't accepted yet, or are there issues with how it was run?
It’s a self-improving harness, and ARC-AGI-3 is explicitly a few-shot benchmark. It’s likely that it gave itself more than the maximum number of tries to learn the games, or even hardcoded the answers.
There’s a lot of improvement to be had from the benchmark harnesses, but sometimes, like with ARC-AGI-3, the limitations are intentional.
Might actually try this