An agent that plans with a frontier model but runs most of tokens locally (built it for my own dual-3090 rig)
A developer has created a new AI agent designed to leverage the planning capabilities of frontier models while executing most tasks with local, less powerful models. This approach aims to balance cost and performance, using a tiered system where a powerful model like Codex handles planning and task decomposition, while local models such as Qwen 3.6 27B perform the bulk of the work. An optional senior model, Kimi K2.6, can be used as a fallback when local models struggle, with the system designed for deterministic validation to prevent context drift and ensure task completion. AI
IMPACT Enables more capable task execution from local models by leveraging frontier models for planning and validation.