Agent and tool-use rankings.

The agent proxy blends available search, document, coding, and latency signals that matter for multi-step workflows. It is useful for orientation, but it is not a full tool-use benchmark until BFCL and Terminal-Bench are wired into the data layer.

BFCL is specifically relevant to function calling and tool use, but the page only includes benchmark families once we have a reliable, auditable feed. Until then, BFCL is treated as a planned source rather than an implied score.

Agents often need to retrieve information, inspect documents, and act on context before producing an answer. Search Arena and Document Arena are proxy signals for those parts of the workflow.

Latency is time to first token. Lower latency can make iterative agent loops feel faster, but it should be balanced against the model's task quality and tool reliability.