How Harvey & Fireworks Beat Closed Source on Cost + Quality

This post details two system-level techniques applied to Harvey's Legal Agent Benchmark (LAB) to improve performance and cost-efficiency: 1) Harness engineering with an open-source GLM 5.1 worker that selectively invokes Claude Opus 4.7 as a callable advisor for complex sub-tasks, achieving better all-pass rates at a lower cost than end-to-end Opus. 2) Fireworks-native post-training (SFT and RFT) of Kimi K2.6 on LAB trajectories, further improving performance and controllability. The post highlights the ability to run both training and serving on the same Fireworks platform, eliminating the research-to-production gap.