$\tau^2$-Bench: Evaluating Conversational Agents in a Dual-Control Environment

Existing benchmarks for conversational AI agents simulate _single-control_ environments, where only the AI agent can use tools to interact with the world, while the user remains a passive information provider. This differs from real-world scenarios like technical support, where users need to actively participate in modifying the state of the (shared) world. In order to address this gap, we introduce $\tau^2$-bench, with four key contributions: (1) A novel Telecom dual-control domain modeled as a Dec-POMDP, where both agent and user make use of tools to act in a shared, dynamic environment that tests both agent coordination and communication; (2) A compositional task generator that programmatically creates diverse, verifiable tasks from atomic components, ensuring domain coverage and controlled complexity; (3) A reliable user simulator tightly coupled with the environment, whose behavior is constrained by tools and observable states, improving simulation fidelity; (4) Fine-grained analysis of agent performance through multiple ablations including separating errors arising from reasoning vs communication/coordination. In particular, our experiments show significant performance drops when agents shift from no-user to dual-control, highlighting the challenges of guiding users. Overall, $\tau^2$-bench provides a controlled testbed for agents that must both reason effectively and guide user actions. Code, data, and leaderboard are available at https://taubench.com/.

Today's tests for AI customer-service assistants place the AI in a world where only it can take actions, while the user just talks. But in real support — like calling about a broken phone — the user must also act, restarting devices or toggling settings, while the agent guides them. To address this gap, we introduce τ²-bench, with four contributions: 1) a new telecom support task where both the AI and a simulated user can take actions on a shared system; 2) an automatic task generator that builds varied, solvable problems from a small set of reusable building blocks; 3) a user simulator whose actions are kept predictable and consistent with what is actually possible by tying it to the environment rather than prompts alone; and 4) measurements that separate the AI's reasoning errors from its communication errors. Experiments show that even the best current assistants succeed on only 34–49% of new tasks, with performance dropping about 20% when the AI must guide a user instead of acting alone. τ²-bench gives researchers a way to measure and improve this gap before AI assistants reach customer-facing roles. Code, data, and a leaderboard are at taubench.com.