Arcade lets agents request tools from a changing catalog
In plain words
Agents can use only the tools they are given, but a long list takes up space in every request and can become outdated. Arcade’s Tool Recommendation system lets an agent describe a task and request tools as needed, instead of receiving the whole catalog.
In 304 of 327 tested Codex sessions, the first request returned every tool the workflow needed. Those sessions covered 30 workflows, so the result describes these tests rather than all agents or tasks.
The research
An agent can use only the tools it has been given. A short fixed list can limit which tasks it can handle, while sending a large list uses more context on every request. How can an agent reach a changing catalog without carrying the whole list each time?
Tool Recommendation lets agents request tools from a changing catalog through a fixed gateway interface, and the study tests its ranking and session costs.
Arcade’s tool catalog changes, but giving an agent a fixed list creates a tradeoff: a small list may miss a task, while a large one uses context on every request. Can an agent reach the needed tools without carrying the whole catalog in each request?
Arcade’s Tool Recommendation system lets agents request tools through a fixed gateway interface. On ToolRet, a test of ranking tools for queries, it placed every needed tool in the top ten for a larger share of queries than BM25 in the same harness.
In paired multi-step workflows, the pooled bill was also lower than with a static list of 100 tools.
These results come from Arcade’s gateway, a ranking test, and a limited set of clients and seeded workflows. They do not establish the same ranking or cost results for other gateways, clients, or workloads.
Analysis
Deep
- Original paper authors
- Sam Partee, River Stevens, Thierry Damiba, Valerie Fanelle
- Submitted to arXiv
- License
- arXiv non-exclusive