How we built Command Code, the most used open-model harness, at 950T tokens
Presented by: Ahmad Awais
Originally aired on October 23 @ 10:00 AM - 10:30 AM EDT
How did Command Code make DeepSeek outperform Claude Opus with tool repair harness engineering research? Now at 950 trillion–token scale and over 100K developers, Command Code is fixing billions of tokens for open models. A couple of months ago, developers did not trust open models. Bounced tool calls, wasted context, and poor cache hit ratios made capable models look broken. We purpose-built a coding agent harness for open models: a repair layer that fixes malformed tool calls, a taste model that learns your coding preferences, and per-provider cache pinning that holds a 99.43% cache hit rate at trillions of tokens per day. In this talk, learn how you can cut your AI costs by 10 to 20 times while also increasing the amount of work done using an abundance of tokens with top open models.
English
