Agent not pulling its weight? First instinct for most devs: throw a more expensive model at it. We’ve been conditioned to believe that bigger price tag equals better results. But what if I told you that swapping out your harness—the orchestration layer around your agents—can turn a bunch of mid-tier models into a beast that outperforms the flagship, at a fraction of the cost?
That’s exactly what the open-source project OpenSquilla just demonstrated in their latest release and technical report. I went through the numbers twice, because they looked too good to be true.
Here’s the headline: compared to running Fable5 (the current state-of-the-art), their new harness achieved a quality score +1.05 higher, while slashing cost by 67.77%. Against Opus 4.8: quality +8.59, cost down 40%. Against GPT-5.5: quality +10.84, cost down 13%.
And what models did they use to pull this off? Not some secret supermodel. Only four open-source Chinese models: DeepSeek v4, GLM-5.2, Kimi K2.7, and Qwen3.7. None of them can beat Fable5 individually. In a one-on-one fight, each of these would lose—badly. But together, orchestrated by a smart harness, they win.
That’s the part most people miss. When your agent fails, you don’t need a better brain. You need a better team structure. The harness decides how agents talk to each other, how they debate, how they vote on answers, and how they recover from failure. OpenSquilla’s secret sauce isn’t a magic model—it’s a multi-agent orchestration pattern that lets weaker models compensate for each other’s blind spots.
Think of it like this: one average engineer won’t beat a genius. But four average engineers with a killer meeting protocol, a good whiteboard, and a system to check each other’s work? They’ll produce better designs than the lone genius, and for way less salary.
The lesson is painful for anyone who’s been blindly upgrading models: your biggest bottleneck is never the model itself, it’s how you’re using it. OpenSquilla’s approach proves that investing in architecture—the harness—delivers more ROI than a more expensive API key. They didn’t invent a new LLM. They just redesigned how multiple LLMs collaborate.
So before you wire up that 100x-cost increase, ask yourself: does your agent really need a smarter brain, or just a better way to work with its teammates? Because a cheap model with a good harness beats an expensive model with a bad one, every time.
And that’s exactly what “putting theory into practice” looks like.