Filter the tool list before inference rather than after: 70% fewer tools shown, 61.7% fewer input tokens, 51% lower latency

arXiv 2608.23078·medium signalXBlueskyLinkedInCopy link

AgentWeave inserts a deterministic pre-inference routing layer that builds a bounded model-visible action space from eligibility, requirement, capability and routing signals, leaving the downstream function-calling model untouched. Against an all-tools baseline it presents 70.18% fewer tools, consumes 61.70% fewer input tokens and halves mean latency. The accuracy claim is deliberately narrow and worth reading skeptically: 6 of 48 fresh BFCL V4 multi-function tasks solved versus 0 of 48 for all-tools, random top-8 and semantic top-8, McNemar p=0.03125, with the authors calling it a routing-pressure study rather than a leaderboard result.