Nine papers covering the tool-use story. API-Bank and Gorilla established the format. ToolBench scaled the API surface. BFCL is the leaderboard most teams actually report against: V3 added the multi-turn dimension, and V4 (July 2025) added web search, memory, and format sensitivity, with agentic tasks weighted at 40% of the overall score. The leaderboard page shows its last update as 2026-04-12 (fetched as of 2026-10-01). TRAJECT-Bench is the trajectory-aware extension. MCP-Atlas (early 2026) runs the harness against 36 real MCP servers rather than mocks; its best pass rate across 20 frontier models was 82.2%, and 63.3% of diagnosed failures were cognitive rather than tool-call related. The 2026 validity audit is the paper to read before you trust a single-run score from BFCL v4 or tau2-bench: across 496 expert-reviewed tasks it found 92 evaluator-human disagreements, and LiveMCPBench scored the same setup anywhere from 57.9% to 76.8% over 23 repeats. AgentDojo lives at the intersection of tool use and security, where prompt injection into a tool call becomes the attack surface. The Yehudai survey closes the picture with the broader agent-eval view.