Feedback and independent evaluation welcome

[SAURAV SINGLA (sauravsingla)]

We recently released AgentWeave: Routing Before Reasoning for Efficient Function Calling in Tool-Rich Language Models.

AgentWeave explores whether reducing the candidate tool/action space before LLM inference can improve function calling while keeping the downstream model unchanged.

I would especially value feedback on the routing methodology, experimental design, baselines, reproducibility, failure cases, limitations, additional benchmarks/models, and comparisons with alternative tool-retrieval approaches.

Critical feedback is very welcome. If anyone independently reproduces the experiments — whether the results agree or disagree with ours — please share the setup, results, and observations.

Paper: https://arxiv.org/abs/2608.23078
Code: https://github.com/sauravsingla/agentweave

The goal is to make AgentWeave more rigorous, reproducible, and useful to the broader tool-calling and agent-systems research community.