# AgentWeave Paper

## AgentWeave: Routing Before Reasoning for Efficient Function Calling in Tool-Rich Language Models

**arXiv:** 2608.23078  
**Paper:** https://arxiv.org/abs/2608.23078

AgentWeave studies deterministic pre-inference routing for tool-rich language-model agents. The paper evaluates whether reducing the model-visible candidate action space before inference can improve function-calling reliability while leaving the downstream model unchanged.

### Citation

**AgentWeave: Routing Before Reasoning for Efficient Function Calling in Tool-Rich Language Models.** arXiv:2608.23078 (2026).

### Software artifact

**Evolving AgentWeave project — Concept DOI:** https://doi.org/10.5281/zenodo.22913459  

**AgentWeave v0.7.0 — Version DOI:** https://doi.org/10.5281/zenodo.22913460  
**Zenodo:** https://zenodo.org/records/22913460  
**GitHub release:** https://github.com/sauravsingla/agentweave/releases/tag/v0.7.0

Use the Concept DOI when referring to AgentWeave as an evolving software project. Use the version DOI when citing or reproducing the archived v0.7.0 software release. The arXiv identifier remains the citation for the research paper.

### Research boundary

The paper reports a frozen BFCL-derived routing-pressure study. It is not an official full BFCL leaderboard submission. The repository keeps frozen evaluation evidence separate from later framework and experimental extensions.