Record metadata

type: "Research"
title: "τ-bench (2024)"
description: "Tool-Agent-User evaluation with domain policies."
resource: "https://arxiv.org/abs/2406.12045"
tags: [benchmark, agents, policy, research]
generated: { by: human:crpage, at: 2026-07-09T09:44:00Z }
status: stable
sources: [{ id: primary, resource: "https://arxiv.org/abs/2406.12045" }]

τ-bench (2024)

Tests reliability and policy adherence under dynamic users with domain policies (airline, retail) — both remain weak. Evidence that policy belongs in infrastructure, not prompts.

Relates to: Policy enforcement.