The description is all the router sees.
An agent has two tools, handbook_search over the Larkspur Cloud handbook and web_search, and twenty mixed questions. Edit either description and every question is routed again. The router here is the M10 TF-IDF retriever pointed at the two descriptions, a stand-in for the model’s choice.
- Routed correctly
- Held out, never tuned on
- Sent to the wrong tool
- Description tokens, sent on every call
Where each question goes
Cosine similarity, TF-IDF over the two descriptions
| Question | Handbook | Web | Routed to | Result |
|---|
A question that matches neither description gets no tool: the agent answers from its own weights. Equal scores above zero count as a miss. Highlighted words are the ones the question shares with a description. Tokens use the four-characters-per-token rule of thumb.