I don't want to be too negative, but ... all this for a 0.8% improvement in SWE-bench Verified (90.2 for OpenCode vs 91)? And why is this (saturated) benchmark the one coding benchmark chosen to showcase on the homepage?
Without trying it, this seems like its probably just a massive waste of tokens.
It's all very glitzy, but I'm failing to understand how "One prompt in. One result out." is of any importance.
This and other recent AI hype-fests all seem to be obsessed with making agents do more work unattended.
But surely in the real world, anyone who's got a real product to make is going to want to steer what's happening. It's ridiculous to think that anyone with a deadline would write a prompt so perfect that they walk away for 4 days and come back to find the finished product ready to ship.
If you really can write a prompt so complete and perfect that it needs nothing further, then any regular harness could probably also do the job. But if like normal people you need to try something, think about it, iterate, and repeat.. then you also just need a regular harness.
> But surely in the real world, anyone who's got a real product to make is going to want to steer what's happening.
The promise of AI is that you won't need to pay people in order to think. A bet that there will be a long term need to steer is also a bet that AI will fail.
The people footing the bills for it are paying because they believe AI will succeed, and there will be no more need for a human in the loop. The reason people are pouring trillions into these AI companies is that they expect we'll make the breakthroughs that make it possible for AI to succeed at automating everything humans are able to do today.
So, skate to where the puck is going, not where it is, and all that.
> But surely in the real world, anyone who's got a real product to make is going to want to steer what's happening.
In my own harness I log each and every user message by hook and use the model to extract user intent by re-reading the chat log from time to time. The raw messages are very important, they contain information that can be used to refine the harness on the one hand, and to validate if the agent still follows user intent on the other. Models tend to get lost in the details and forget the big picture.
Not saying you're wrong. But right now I'm trying to craft some skill prose to instruct an agent how to optimize a certain process based on my own heuristics, and it's failing. Maybe I can let a super agent divine the right skill prose, iterating to see what works. It's worth a try.
OK, but by the time you've worked out the correct prompt to tell your meta-agent how to iterate on optimising the right prompt for your actual agent, perhaps you could have just tried a few things and got it going!
RSI here is for "recursive self-improvement", instead of a harness being built to help users with repetitive strain injury like I first thought when reading the post title
Well for starters the use of the term "Recursive" is very dubious.
This is as far as the eye can see all very iterative, there is no tail-call or anything fancy, it is loops. Also "Recursive" I feel kind of tries to imply the LLM's weights are being pushed around in some feedback, because to recurse you have to invoke the thing you're recursing into at its very start right? That would be reinforcement probably, and there is none of that in any such RSI so far, at least not public. Please someone contradict me with examples.
So please: "ISI" for iterative self improvement is fine. Also "ISI" does not fit the (outdated!) Vernon Vinge "singularity" trope, and that is a good thing!
The example in the docs of improving nanochat is iterative. It's a looped process in one thing altering a second thing.
What would be recursive is raven updating raven to make it better at doing things. For what I picture as RSI the important part would be that it's able to make itself better at doing things and better at improving itself.
Now there's an "evolver" part that improves the harness over time but I don't know how far that goes or what scope it has to update things.
I guess it's that if you have a function called "optimise" that takes functions and makes them better, calling optimise(my_process) is iterative regardless of how many times you do it. Calling optimise(optimise) is inherently different.
A register is the obvious choice, but of course then you get a gold rush and people squatting on XPZ and stuff without even having invented an opaque term behind it.
A very popular yoga kata called sun salutations (available in various versions depending on your fitness/advancement level) helped me get rid of wrist and thumb RSI. It stretches and gently stresses these load bearing ;) joints, thus strengthening them.
Hope this helps the last few folk before searching for RSI will be impossible due to the term being taken over.
For those afraid of Satanism in yoga, there's also non satanic variants where you don't greet each other saying Namaste or say Shanti anywhere during the practice.
As long as we’re burying things for archaeologists to find, let it be known that AGI once stood for “adjusted gross income”, a measure used for income taxes.
I continue to yearn for a harness of harnesses but each one I try (or build) takes me uncomfortably far from the work being done.
I don't want to be a prompt shuttle, though I feel that way sometimes. Performing the same dance for each ticket I work on. My issue is that, to bastardize a common joke/phrase, 50% of the things the agent stops for are things it (or another agent) could answer for me, but it's a different 50% task to task.
With HoH's I constantly feel like I'm getting peppered with unimportant questions or being kept out of the loop of things that really need my eyes on it. Threading that needle has been particularly difficult.
An accusation wrapped as a genuine question. Bravo.
Let me turn this around to you. Do you think your timelines on Reddit/X are indicative of the tech scene of {London,China,India,Indonesia}? What makes you think you know of every popular project out there?
Reminds me a lot of omnigent (which I am a huge fan of) with a persistent memory layer. Unlike omnigent's subagent threads, the DAG it uses to coordinate other harnesses doesn't look to be durable; I am curious as to whether this is by design or is a forthcoming feature, as this essentially makes or breaks my use case of long-running project-sized implementation sessions.
In any event, it's great to see competition in this meta-harness space, which is likely one that none of the frontier labs will touch since it, by definition, would utilize their competitors' products.
I'm curious if you have a few mins for feedback, what are the top 2 things here that omnigent does that is significantly better for you than latest cc/codex which can launch subagents, auto save memory of a project.
I'm wondering that as these core tools continue to enhance and add these capabilities, how much of a benefit these meta harnesses actually provide.
This looks similar to https://paseo.sh/, if I understand correctly. I’ve recently tried it and liked it a lot. Would be nice to see a comparison. When’s the harness of harnesses of harnesses coming?
I'm using paseo heavily too. The main differences here seem to be around how opinionated Raven's orchestration is. They're providing agents, workflows, memory, skills, etc. Paseo gives you some orchestration tools but it's mostly letting the 'native' harnesses do the work. For my workflow, I'm interested in some of what they're doing here but I'm not in buying into their whole system (markdown for memory and calling it RSI, as someone else called out, is not doing it for me). The follow up actions and meta-harness tuning look pretty cool.
They also don't ship an app which is one of the best parts of paseo. Then again Paseo's perf leaves a lot to be desired.
Yeah I actually don’t even use Paseo’s orchestration much. I’m mostly using Paseo because I want to use Anthropic and OpenAI’s native harnesses for their respective models, but OpenCode for others, with a good remote mobile app experience when I want to check in on agent work or steer things away from my computer.
I haven’t found a better OpenCode remote mobile app experience. If OpenCode or Pi makes one, I might just move to that.
I always figured that the free/open weights models like qwen3.8:27b would perform just as well if not better than Claude's latest if you just fed it back into itself enough times. This project seems to prove that this is indeed the case.
What I'd like to see now is how good it can get when you feed the micro models like qwen3.5:0.8b into itself to solve problems. Will it be like toddlers discussing neighborhood politics at a pretend tea party or will it actually get some decent results?
Another game-changer (if this style works out): Just get a model like qwen3.8:27b onto one of those model-on-a-chip cards that makes it 1000x faster and see how fast it can go using the same method.
What you describe reminds me of some studies of jumping spiders. Some aspects of their intelligence (like counting) matches that of a 1 year old human, but because their brains are so tiny it just takes them much longer to do the same counting. IOW it's not the size of the model, but how it's organized and the strategies for using it.
This article has lots of fluff but it describes a lot of what I'm talking about:
Harnesses are only useful when frontier models are incapable of designing their own efficient interfaces with systems which they are improving at rapidly. This will be seen as a transitional artifact of a specific time in the development of general intellegent systems
Interesting. Only thing, from someone who has built something similar, is that it moght tend to duplicate certain capabilities that those harnesses handle on their own. Some overlap is bound to happen.
This, and also they have to chase changes in model tuning and capabilities. I have a nice little meta-harness focused on product development (requiremwnts, acceptance criteria) for Claude code, and when major new models come it it’s weeks before I can find and fix constraints that are no longer necessary + constraints that have become necessary.
I think the only way to do that is test yourself... otherwise, you are asking for bias. DSH's "cordis" is very interesting, I haven't tried it yet. I have spent lots of time with pi.dev, omp and hermes. Aspects of all harnesses are great, but there is always something that bugs me.
If you have the chops to evaluate different harnesses, you have the chops to build one that is perfect for you.
Without trying it, this seems like its probably just a massive waste of tokens.
This and other recent AI hype-fests all seem to be obsessed with making agents do more work unattended.
But surely in the real world, anyone who's got a real product to make is going to want to steer what's happening. It's ridiculous to think that anyone with a deadline would write a prompt so perfect that they walk away for 4 days and come back to find the finished product ready to ship.
If you really can write a prompt so complete and perfect that it needs nothing further, then any regular harness could probably also do the job. But if like normal people you need to try something, think about it, iterate, and repeat.. then you also just need a regular harness.
The promise of AI is that you won't need to pay people in order to think. A bet that there will be a long term need to steer is also a bet that AI will fail.
The people footing the bills for it are paying because they believe AI will succeed, and there will be no more need for a human in the loop. The reason people are pouring trillions into these AI companies is that they expect we'll make the breakthroughs that make it possible for AI to succeed at automating everything humans are able to do today.
So, skate to where the puck is going, not where it is, and all that.
In my own harness I log each and every user message by hook and use the model to extract user intent by re-reading the chat log from time to time. The raw messages are very important, they contain information that can be used to refine the harness on the one hand, and to validate if the agent still follows user intent on the other. Models tend to get lost in the details and forget the big picture.
Or put another way: "You can't fool me, it's turtles all the way down!"
This is as far as the eye can see all very iterative, there is no tail-call or anything fancy, it is loops. Also "Recursive" I feel kind of tries to imply the LLM's weights are being pushed around in some feedback, because to recurse you have to invoke the thing you're recursing into at its very start right? That would be reinforcement probably, and there is none of that in any such RSI so far, at least not public. Please someone contradict me with examples.
So please: "ISI" for iterative self improvement is fine. Also "ISI" does not fit the (outdated!) Vernon Vinge "singularity" trope, and that is a good thing!
The example in the docs of improving nanochat is iterative. It's a looped process in one thing altering a second thing.
What would be recursive is raven updating raven to make it better at doing things. For what I picture as RSI the important part would be that it's able to make itself better at doing things and better at improving itself.
Now there's an "evolver" part that improves the harness over time but I don't know how far that goes or what scope it has to update things.
I guess it's that if you have a function called "optimise" that takes functions and makes them better, calling optimise(my_process) is iterative regardless of how many times you do it. Calling optimise(optimise) is inherently different.
Hope this helps the last few folk before searching for RSI will be impossible due to the term being taken over.
For those afraid of Satanism in yoga, there's also non satanic variants where you don't greet each other saying Namaste or say Shanti anywhere during the practice.
I don't want to be a prompt shuttle, though I feel that way sometimes. Performing the same dance for each ticket I work on. My issue is that, to bastardize a common joke/phrase, 50% of the things the agent stops for are things it (or another agent) could answer for me, but it's a different 50% task to task.
With HoH's I constantly feel like I'm getting peppered with unimportant questions or being kept out of the loop of things that really need my eyes on it. Threading that needle has been particularly difficult.
Let me turn this around to you. Do you think your timelines on Reddit/X are indicative of the tech scene of {London,China,India,Indonesia}? What makes you think you know of every popular project out there?
In any event, it's great to see competition in this meta-harness space, which is likely one that none of the frontier labs will touch since it, by definition, would utilize their competitors' products.
I'm wondering that as these core tools continue to enhance and add these capabilities, how much of a benefit these meta harnesses actually provide.
They also don't ship an app which is one of the best parts of paseo. Then again Paseo's perf leaves a lot to be desired.
I haven’t found a better OpenCode remote mobile app experience. If OpenCode or Pi makes one, I might just move to that.
What I'd like to see now is how good it can get when you feed the micro models like qwen3.5:0.8b into itself to solve problems. Will it be like toddlers discussing neighborhood politics at a pretend tea party or will it actually get some decent results?
Another game-changer (if this style works out): Just get a model like qwen3.8:27b onto one of those model-on-a-chip cards that makes it 1000x faster and see how fast it can go using the same method.
This article has lots of fluff but it describes a lot of what I'm talking about:
https://knowablemagazine.org/content/article/mind/2021/are-s...
If you have the chops to evaluate different harnesses, you have the chops to build one that is perfect for you.