AgentWeave — pre-inference routing for tool-rich agent systems

[SAURAV SINGLA (sauravsingla)]

Sharing AgentWeave, an open-source orchestration framework exploring pre-inference capability/tool routing for heterogeneous agent systems.

The core idea is:

do not ask the LLM to reason over every available capability if a deterministic routing layer can first construct a smaller task-relevant action space.

For function calling, the architecture is:

task
→ candidate tools
→ AgentWeave routing
→ bounded tool subset
→ underlying function-calling model
→ native evaluator

A recent frozen BFCL-derived routing-pressure study using Hammer2.1-1.5B produced:

• AgentWeave: 6/48 native successes
• all-tools: 0/48
• random top-8: 0/48
• semantic top-8: 0/48

with substantially fewer model-visible tools and input tokens.

These are research results under a custom routing-pressure setup, not official BFCL leaderboard scores.

We’ve now opened a BFCL-compatible upstream PR to enable standard evaluation:
ShishirPatil/gorilla#1358

AgentWeave:
https://github.com/sauravsingla/agentweave

I’d especially value feedback from people working on large tool catalogs, dynamic tool selection, and multi-agent orchestration.