Agents & Tool Use
WorkBench Revisited
WorkBench Revisited (2026) — workplace-agent benchmark of 690 sandbox tasks (5 databases, 26 tools), graded by comparing the sandbox final state to ground truth. Per-(model, task) binary task-completion outcomes for ~24 models (2023-2026) run under a native tool-calling harness, reproduced from the released per-task agent runs via the repo scoring pipeline.
690items
24subjects
MITlicense
agents_and_tool_usedomain
textmodality
item-level responses released
Saturation status: Yes
Response matrix
Loading response matrix…
Correct (1)Incorrect (0)Unobserved
Scale: 1 = correct · 0 = incorrect