Summary
The benchmark evaluates 12 agents on questions involving hidden facts distributed across generated company systems. The strongest agent answered 18 of 24 hidden-knowledge attempts correctly, while all agents together solved only two of 84 hardest attempts.
Why it matters
Enterprise deployment depends on extracting implicit information across systems, a capability that ordinary retrieval and straightforward tool-use benchmarks may overstate.
Related coverage
- AgentX demonstrates long-horizon autonomous model research
- Bold Penguin launches DeX ai for insurance-marketplace operations
- webAI secures a $30 million enterprise deployment commitment