This looks very similar to what I've seen with a different harness - https://github.com/AMAP-ML/LongHorizon-Harness. It's critique pattern looks like another version of lh-harness without looping toward an overarching goal
> One model drafts a result, an independent read-only critic from a different model family reviews it
Multiple model vendors is key here, the cascade pattern doesn't need it, but the critique pattern does.
Last Nov, my team wrote a paper ("Team of Rivals") on the difference between using an OpenAI model to Critique an Anthropic model's output vs running a self-review agent loop on the same vendor.
The ablations [1] proved that neither company alone was better than using both.
The paper was a general response to "What does your company do that Anthropic can't?" but more so a demonstration of how to make something 90%+ good with models which eval at 60% or so (& Gas Town post unblocked our "this is a trade secret" argument about the paper).
My editor and gatekeeper use like the same model Opus 5. Different prompts and a kind of different input data. The gatekeeper receives the fact check results next to finished text. In the same time the editor already delivered them.
As far as I remember over the entire period he removed 27 posts out of 187 that went through him.
So I believe that different manufacturers are not mandatory. What matters I guess is not the difference in models but... the fact that the critic has a different input and doesn’t have their own text that needs to be defended. That's the point
I'm questioning their results. It doesn't take much to beat the frontier in single benchmarks if one puts extra software between the model and the harness.
This is also a reason why comparing "naked models" for which weights are available and frontier where providers can do whatever they want behind the scenes is unfair.
Specifically, all it took to boost Qwen3.8-27B to get 10% more points on SWEbench Pro and Terminal Bench 2.0 with a proxy that has just these basics:
- tweaks few decode settings like slightly higher temperature
- detects when model gets stuck and tells it to "go on"
- detects responses cut in the middle, empty responses that contain only reasoning, formats not passing verification etc and tells the model to "try better"
And that is it. 10% more. I admit on a subset of tasks, but results are results, even on a subset.
Yeah but what is the point of "naked model" evaluations? It seems to me that the only interesting question is capability per time and cost (and conditioned on other important things, like privacy). It doesn't matter to me which part of that equation is being implemented by model weights and which part by other supporting software.
Harnesses can fairly easily be adapted to other models. Moving capabilities from one model to another is a more involved process requiring distillation, training, etc.
Why? I feel it kinda implies what the approach is. Could definitely be worse, strongly prefer this over just giving the thing a random-ass name like Laguna.
i've been using adversarial critique and reviews for many planning, solution design and implementation steps inside workflows.
It's so effective and helps catching so many design flaws, implementations misses etc ... that i'm wondering how people manage to build complex/large projects with agents without this kind of process. Well, i actually built this thing because i couldn't get good results so i had to find a way.
I'm gonna open source the whole thing but it needs some cleanup, there's a basic landing page here https://kodfactory.com if anyone wants to be notified when it's released on github. Yeah i know, the world really needs another software factory :-)
I have a custom agent that I call my "G.S.D." (Get Stuff Done) agent.
It is explicitly not a foreman, task routing, or an orchestrator agent. It has a bias toward direct action and is instructed to only delegate when necessary.
I've found that this approach yields significantly faster results, without much of a quality trade-off, than an agent whose primary impulse is to delegate.
In contrast, HydraFusion starts with a task routing step, then sequential planning, execution, and review stages. My guess is that this workflow is best for people who are prioritizing cost over speed for the same level of quality.
> In controlled offline evaluations, HydraFusion’s selective coding workflows matched or exceeded the evaluated Opus 5 baseline
Opus 5 (in practice) is not a good baseline to compare against.
I'm shocked that they chose to primarily compare against Opus 5 in all the article's charts. It's pretty disingenuous that they're claiming "frontier" quality, but didn't compare against Fable or Sol.
I would imagine they did not test against Fable because Microsoft and GitHub (like many big companies) have internally given the instruction not to use this model, because of the data retention policy.
They did also compare to Sol and the comparisons are still favorable. However, they were most favorable comparing to Opus 5 because of the cost judging by the charts.
Multiple model vendors is key here, the cascade pattern doesn't need it, but the critique pattern does.
Last Nov, my team wrote a paper ("Team of Rivals") on the difference between using an OpenAI model to Critique an Anthropic model's output vs running a self-review agent loop on the same vendor.
The ablations [1] proved that neither company alone was better than using both.
The paper was a general response to "What does your company do that Anthropic can't?" but more so a demonstration of how to make something 90%+ good with models which eval at 60% or so (& Gas Town post unblocked our "this is a trade secret" argument about the paper).
[1] - https://github.com/t3rmin4t0r/critique-evals
This is also a reason why comparing "naked models" for which weights are available and frontier where providers can do whatever they want behind the scenes is unfair.
Specifically, all it took to boost Qwen3.8-27B to get 10% more points on SWEbench Pro and Terminal Bench 2.0 with a proxy that has just these basics: - tweaks few decode settings like slightly higher temperature - detects when model gets stuck and tells it to "go on" - detects responses cut in the middle, empty responses that contain only reasoning, formats not passing verification etc and tells the model to "try better"
And that is it. 10% more. I admit on a subset of tasks, but results are results, even on a subset.
Do you have a link/paper for this you could share?
HyDRA: Hybrid Dynamic Routing Architecture for Heterogeneous LLM Pools : https://arxiv.org/pdf/2605.17106
But could be your thing too obviously.
It's so effective and helps catching so many design flaws, implementations misses etc ... that i'm wondering how people manage to build complex/large projects with agents without this kind of process. Well, i actually built this thing because i couldn't get good results so i had to find a way.
I'm gonna open source the whole thing but it needs some cleanup, there's a basic landing page here https://kodfactory.com if anyone wants to be notified when it's released on github. Yeah i know, the world really needs another software factory :-)
It is explicitly not a foreman, task routing, or an orchestrator agent. It has a bias toward direct action and is instructed to only delegate when necessary.
I've found that this approach yields significantly faster results, without much of a quality trade-off, than an agent whose primary impulse is to delegate.
In contrast, HydraFusion starts with a task routing step, then sequential planning, execution, and review stages. My guess is that this workflow is best for people who are prioritizing cost over speed for the same level of quality.
> In controlled offline evaluations, HydraFusion’s selective coding workflows matched or exceeded the evaluated Opus 5 baseline
Opus 5 (in practice) is not a good baseline to compare against.
I'm shocked that they chose to primarily compare against Opus 5 in all the article's charts. It's pretty disingenuous that they're claiming "frontier" quality, but didn't compare against Fable or Sol.