- Microsoft's Julia Kasper said model providers sometimes send GitHub advice on a model's habits, after GitHub shares its own test findings, and GitHub tunes its harness.
- Our reading: a buyer gets a model plus a harness tuned around it. Kasper did not say whether that tuning is published.
Julia Kasper, a product manager at Microsoft, said model makers sometimes advise the GitHub Copilot team on a model's habits. GitHub then tunes its own software around each model. She spoke on GitHub's show about how Copilot handles every new model.
What was said
Kasper said GitHub holds weekly check-ins with the big model providers. Their contracts let providers share a little in advance when a new model may arrive. GitHub adds the model to its platform and starts tuning its harness. The harness is the software around the model that gives it instructions and tools.
That work starts before launch. GitHub runs offline tests and uses the model inside the company. It shares what it finds with the provider. Sometimes the provider then sends back an updated checkpoint, which is a new trained version. Kasper said GitHub also gets recommendations back, because providers are closer to model training.
Her example was a model's personality. Providers know it, she said: "Maybe they're a little bit more or less lazy to call a certain tool." GitHub uses such advice to fine-tune the harness.
Harness choices matter, she said: "So the same model, a different harness could essentially heavily influence the quality of an outcome." She also said Copilot launched 25 new models in one quarter.
Why it matters
Here is WebPulse's explanation of the mechanism, not Kasper's. A coding agent works when the model chooses to call tools, such as a file search or a test run. A model that is slow to do so may guess when it could have checked. A harness can respond through its instructions and how it describes its tools. Kasper did not say what GitHub changes.
Our reading: this is why the same model can behave differently on different platforms. A buyer comparing two tools may be comparing two sets of tuning, not just two models. Kasper did not say whether any of that tuning is published, so buyers can ask which model habits a vendor tuned around.
The other side
Kasper framed the arrangement as a win for both sides, and said working with different vendors keeps GitHub honest. She did not say users are kept in the dark, and she pointed to a benchmark site, benchmark.github.com. The excerpt ends before she says what it shows. The talk also left open what else providers recommend, and how far a harness change can hide a weakness in the model.
Written by the WebPulse Newsroom with AI assistance, and checked by our editorial review: every quotation was verified against the recording's transcript. How we use AI.
The conversation this talking point comes from
- GitHub: How we optimize GitHub Copilot for every new model (2026-11-18)





