{ "benchmark": "tau-bench Retail", "trajectories_per_agent": 2280, "tasks": 228, "agents": [ { "model": "gpt-5-nano-2025-08-07", "reasoning_effort": "minimal", "max_completion_tokens": 4000 }, { "model": "claude-haiku-5-5", "thinking": "disabled", "effort": "low", "max_completion_tokens": 4000 } ], "reviewer": { "model": "gpt-6-luna", "reasoning_effort": "high", "max_completion_tokens": 6000 }, "user_simulator": { "model": "gpt-4.1-mini-2025-04-14", "temperature": 0.0, "max_completion_tokens": 2048 }, "assignment_probability": 0.5, "max_steps": 100, "control": "Execute the original frozen tool proposal or batch", "intervention": "Approve, replace with one tool call, or ask one customer-facing question", "invalid_delivery": "Execute the original proposal and retain assignment", "outcome": "Terminal task success under the paper scoring rules" }