README.md

---
language:
- en
tags:
- agent
pretty_name: TPS-Bench
size_categories:
- n<1K
configs:
- config_name: TPS-EASY
  data_files: "TPS-EASY.jsonl"
- config_name: TPS-HARD
  data_files: "TPS-HARD.jsonl"
---

# TPS-Bench: Evaluating AI Agents’ Tool Planning & Scheduling Abilities in Compounding Tasks

[![Paper](https://img.shields.io/badge/arXiv-2511.01527-b31b1b.svg)](https://arxiv.org/abs/2511.01527)
[![GitHub](https://img.shields.io/badge/GitHub-Repository-black.svg)](https://github.com/hanwenxu1/mcp-agent)

## Dataset Description

**TPS-Bench** is a benchmark designed to evaluate the tool planning and scheduling capabilities of AI agents in compounding tasks. Unlike standard tool-use benchmarks that focus primarily on simple sequential execution, TPS-Bench assesses an agent's ability to handle task dependencies, concurrent execution, and complex temporal constraints.

### Subsets

Currently, the benchmark releases two data subsets:
* **TPS-EASY**: Contains 100 instances. Features shorter dependency paths, suitable for evaluating foundational planning capabilities.
* **TPS-HARD**: Contains 100 instances. Involves highly complex, long-chain dependencies and multi-tool parallel scheduling scenarios, requiring advanced logical reasoning and precise temporal execution.