TechPrufer
Sign inGet started

Blogs›Fireworks AI›Hybrid Worker-Advisor AI Architecture

Hybrid Worker-Advisor AI Architecture

Hybrid Worker-Advisor AI Architecture

2
posts
2026

Introduced a hybrid AI architecture combining open-source worker models with closed-source advisor models. This approach leverages the cost-efficiency of open-source models for task execution while utilizing frontier models for targeted review and quality assurance, demonstrating improved performance across various benchmarks like SWE-bench Pro, Terminal-Bench 2.1, and the Legal Agent Benchmark. This post details the application of this architecture to the Legal Agent Benchmark, showing how a GLM 5.1 worker can selectively invoke Claude Opus 4.7 as an advisor to improve outcomes and reduce costs compared to end-to-end frontier model usage. Additionally, it explores the impact of Fireworks-native post-training (SFT and RFT) on Kimi K2.6 for improving performance on the Legal Agent Benchmark.

2026

Frontier AI at a fraction of the cost: open-source worker agents with a closed-source advisor.

6/23/2026

This post details the research and implementation of a hybrid AI architecture. It describes the 'worker-advisor' setup where an open-source model (Kimi-K2.6 or GLM-5.2) performs task execution and a closed-source frontier model (Claude Opus 4.8) acts as a read-only advisor during a review step. The post presents benchmark results showing performance lifts and cost reductions compared to using only open-source or only frontier models. It also includes ablation studies on advisor call frequency and model roles.

How Harvey & Fireworks Beat Closed Source on Cost + Quality

6/3/2026

This post details the application of the hybrid worker-advisor AI architecture to the Legal Agent Benchmark. It demonstrates how a GLM 5.1 worker can selectively invoke Claude Opus 4.7 as an advisor for specific sub-tasks, achieving a higher all-pass rate (18/100) at a lower cost ($368) compared to using Claude Opus 4.7 end-to-end ($954). It also details the results of supervised fine-tuning (SFT) and reinforcement fine-tuning (RFT) of Kimi K2.6 on Legal Agent Benchmark trajectories, showing improvements in all-pass rate and mean score. The post emphasizes the integration of these techniques on the Fireworks platform for seamless experimentation and production deployment.