Context
Arcade.dev researchers describe a gateway that exposes an approximately 8,200-tool catalog through three agent-facing interfaces. On the public ToolRet benchmark, their recommendation pipeline improves complete tool-set retrieval relative to BM25; in seeded workflows, they report lower billed cost than a static 100-tool list.
What changed
The October 16, 2026 original preprint provides a documented deployed-system design, public benchmark comparison and paired synthetic workflow-cost evaluation for tool recommendation in multi-step agents. The research was missing from the corpus.
Why it matters
Agent discoverability requires retrieving the entire set of tools needed to complete a task, not just some individually relevant APIs. Catalog scale, authorization, reranking and token cost jointly shape which actions an agent can discover and perform.
Visibility OS interpretation
Visibility OS interpretation: whole-tool-set completeness and task-decomposition misses deserve separate agent-discovery metrics, while session cost should be evaluated in billed dollars rather than token counts alone. These findings are not evidence of commercial AI web-search rankings or citations.
What remains uncertain
Can independent replications validate the vendor-authored results? How do results change with real customer permissions, larger static baselines, and different model caching? Do improvements transfer from ToolRet ranking to production task completion?
Evidence claims
On ToolRet (7,961 queries, 44,453 tools), Arcade.dev researchers report Completeness@10 of 45.2% for Tool Recommendation versus 28.8% for BM25 within their own evaluation harness (57% relative improvement).
Arcade.dev researchers describe a gateway exposing roughly 8,200 catalog tools through three agent-facing interfaces, while filtering recommendations by eligible tool scope and user authorization.
In Arcade.dev's paired seeded-workflow evaluation against a static 100-tool list, the authors report 13.6% lower total billed cost overall across Codex and Claude Code clients, with no statistically significant difference in task completion.
Related historical records
Multi-tool agent retrieval study finds complete tool-set discovery is harder than single-tool matching
This record related to the linked event.