χ-Bench · NeurIPS 2026
Can agents finish policy-rich healthcare workflows end to end?
75 long-horizon tasks across provider, payer and care management work, in a simulator of 21 healthcare applications with 200+ tools and a 1,279-document policy handbook.
- Best score among thirty frontier agents
- 28%
- Every agent's pass rate when the same task must succeed three runs in a row
- <20%









