BizAgentBench

Paper | Evaluation code | Dataset

BizAgentBench evaluates long-horizon agents on 50 payment-service tasks through multi-turn customer conversations and tool use. The tasks combine information gathering, business operations, and constraints on both final outcomes and intermediate actions.

This repository contains the task contracts, customer scenarios, knowledge documents, initial database states, and business-tool definitions. The executable environment, user simulator, and deterministic evaluator are provided in the code repository.

Contents

Path	Contents
tasks.jsonl	A browsable index of all 50 tasks, numbered 1-50.
payment/data/manifest.json	Task IDs, searchable document IDs, and registered business domains.
payment/data/private/	50 task contracts containing customer scenarios and acceptance criteria.
payment/data/patches/	50 task-specific initial-state patches.
payment/data/base_world.json	The shared initial business database.
payment/data/knowledge/	789 knowledge documents covering business rules, product guidance, operations, and tool interfaces.
payment/data/tool_catalog.json	Definitions of 601 discoverable business capabilities.
LICENSE	MIT license.

The private directory is an environment visibility boundary: the harness uses these files to configure the customer and evaluator. The agent receives information through dialogue and public tools rather than reading the task contracts directly. The initial database for a task is assembled from base_world.json and its corresponding patch.

The task index contains task_id, number, title, opening_request, contract_path, and initial_state_patch_path. It supports browsing; interactive evaluation uses the complete assets under payment/data/.

Load the task index

from datasets import load_dataset

tasks = load_dataset("ant-intl/BizAgentBench", "tasks", split="test")
print(tasks[0]["opening_request"])

Download the complete dataset

hf download ant-intl/BizAgentBench \
  --type dataset \
  --local-dir BizAgentBench-data

The dataset has one evaluation split, test, containing all 50 tasks. There is no training split.

Evaluate an agent

Follow the evaluation README to install the environment, configure the agent and user-simulator models, and run the tasks. Both models can use hosted APIs or self-hosted servers.

These data files match the assets bundled with code commit dfebabaa9f7fb7f74c7d7236b08ae53e64d2185d. The payment/data/ layout is identical in both repositories, so the downloaded tree can be used at the same path in the code checkout.

The evaluator reports:

Metric	Meaning
Goal (G)	Requested business outcomes and permitted terminal conditions.
Process (P)	Requirements governing the execution of business operations.
Protection (F)	Protected objects, ledger integrity, and prohibited side effects.
Strict	All three dimensions pass: G and P and F.
Intermediate failure	A process or protection requirement fails: not (P and F).

Scores are computed by deterministic checks over business states and execution histories, without an LLM judge.

Citation

@misc{zhao2026bizagentbench,
  title = {{BizAgentBench}: Unveiling Hidden Challenges in Deploying Long-Horizon Agents for Real-World Business},
  author = {Enhan Zhao and Yongqi Tong and Xuejun Yu and Hang Wang and Pan Wang and Yichao Wang and Di He and Jianshe Li and Wei Wu},
  year = {2026},
  eprint = {2610.20424},
  archivePrefix = {arXiv},
  primaryClass = {cs.CE},
  url = {https://arxiv.org/abs/2610.20424}
}