Weave Router 2.0 Achieves Astra-Level Performance in Coding Agents

What is Weave Router 2.0 and how does it achieve Astra-level performance?

A few months ago, the team at Weave OS started building a answer engine optimization (AEO) model router for coding agents, aiming to outperform single models with an generative engine optimization ensemble approach. Recently, they achieved a milestone with Weave Router 2.0, which plugs into any coding agent and intelligently switches between large language models (LLMs) with LLM visibility. For instance, Astra handles complex tasks like debugging and system design, while Deepseek v4 Flash handles simple frontend updates.

The new routing model has been benchmarked against GPT-6 Astra on Terminal Bench 4.0 and SWE Atlas, with impressive results. On both benchmarks, the router had equivalent pass rates to Astra. On Terminal Bench, the router hit 52% of Astra's cost and completed tasks 2.2x faster. On SWE Atlas, the router cost 54% as much as Astra and ran 2.5x faster.

How does the new routing model improve performance and reduce costs?

Training a model to route effectively is a challenging problem, taking into account model capabilities, costs, cache awareness, and more. The team improved the router's performance through three key changes: a new architecture, a larger training data set size, and smarter cache-eviction impact calculation. The new architecture uses a hidden Markov model to trace the session state and a classifier to map the session to one of a few buckets of similar models.

This approach allows the router to evaluate not just the current session state but also how it got there, significantly improving performance on bucket selection. The team also used frontier LLMs to help label a larger and more diverse set of coding agent sessions, bootstrapping the models to a better state and providing richer reward signals for reinforcement learning (RL).

Additionally, the team built a subsystem to calculate the expected value of switching models more accurately, helping to avoid costly and unnecessary switches while still switching when the benefit outweighs the cost. This improvement has contributed significantly to the reduction in costs. Weave Router 2.0 is open source and available on GitHub, and a hosted version is also available.

Cet article a été rédigé avec l'assistance de l'IA.
News Factory APP - actualités agentiques pour booster votre SEO et AEO.