AICodeKing spent time putting Command Code through its paces with open models, specifically testing whether the harness — not the model — is what actually determines whether an agent feels competent or falls apart.
The thesis: harness over model
The core argument of the video is blunt.
"The harness is the real story here, not the models. If the harness is bad, even the best model in the world will flail around, break your code, and burn your tokens. If the harness is good, even a mid model can look surprisingly competent."
That's not a defense of weaker models. It's a claim about where the real bottleneck sits in most coding agents.
Tool calling was the test case
A common line against open models is that they can't reliably call tools. AICodeKing's take is that this is a harness problem, not a model problem.
"Command Code's whole thesis is that when people say open models can't tool call, the models aren't the problem. Your harness is... Command Code's harness is designed from the start to make tool calling reliable on open models. I've run open models in other agents, and the same model just performs worse when the harness keeps mishandling its tool calls."
Watching it debug
The video walks through Command Code recovering from a failing test run without spiraling.
"This is where open models usually fall apart in other agents. It reads the test output, finds the actual cause, fixes it, and reruns. No flailing, no hallucinated tool calls, no getting stuck in a loop. GLM 5.2 inside this harness gets you a real chunk of frontier quality for a tiny fraction of the price."
Why this matters
Most comparisons of open vs. closed models are really comparisons of open-model-in-a-bad-harness vs. closed-model-in-a-good-harness — which tells you nothing about the model itself. AICodeKing's video isolates the variable: same model, different harness, very different outcome. That's the bet Command Code is built on.
"The gap between open and closed models is smaller than the gap between good and bad harnesses. That's the biggest takeaway."
Try it
1npm i -g command-codeRun cmd, pick GLM 5.2 or another open model, and see how far a good harness takes it.
