Praex
The detail that gets me is "runs on its own AI models." Every other coding agent out there is essentially a really good wrapper around someone else's frontier model. Praex is saying we build the brain too, which is a much heavier lift but also a much more defensible position if it works.
I keep thinking about where this fits in the broader trend. We've seen companies like Databricks and Snowflake go vertical on data infrastructure. Now the dev-tools layer is starting to do the same with model infrastructure. The question is whether a small team can actually compete on model quality while also building the agent experience. That's two full-time research problems.
I'd want to know where these models are trained, what data they're using, and how they measure up on real-world repo tasks, not just synthetic benchmarks. The cost structure matters too. Running your own inference is expensive, and that pricing pressure could be the real story here. If they can make the economics work while others pay per-token to the big labs, that's a real wedge.