JesseLiu/agentorch-benchmark

AgentOrch benchmark

Model-visible prompts for the AgentOrch benchmark from the paper AgentOrch: How Close Can Global Orchestration Come to Step-by-Step Interaction?. AgentOrch compares global orchestration (the model writes one complete executable workflow before execution) with step-by-step interaction (the model picks one tool call at a time from execution feedback) on the same 650 enterprise standard operating procedures (SOPs) and 3,220 runtime variants, under matched tools, simulated execution environment and evaluation.

The data/ configs hold only what a model sees. The evaluator-side fixtures (reference workflows, reference execution traces, hidden connector state) are in source/ as archives, and the simulator and scorer live in the code repository: https://github.com/jesseliu0913/AgentOrch

Configurations

config	rows	one row per	use
global	650	SOP	Global orchestration: generate one complete workflow from the task specification; it is reused across every runtime variant of the SOP
local	650	SOP	Incremental construction with an append-only editor and validator feedback (no runtime execution)
react	3,220	runtime variant	Step-by-step interaction: one tool call per turn with the concrete runtime input and intermediate results
replan	3,220	runtime variant	Execution-feedback revision of a global workflow

Fields

• benchmark_schema: schema identifier of the export.
• mode: one of global, local, react, replan.
• recipe_id: the source SOP identifier; the same id links the four configs and the ground truth in the code repository.
• variant_id (react and replan only): the runtime variant identifier.
• task: the model-visible prompt text, including the task description, public input/output schemas, execution context, workspace resources, communication templates and the 64 available tool interfaces.

Evaluation

Generated workflows and trajectories are executed in the deterministic connector simulator shipped with the code repository and scored against the reference executions (Orchestration Accuracy, Final Accuracy, Side-Effect Accuracy, Execution-Trace Accuracy, Terminal-Output Accuracy, Error Rate). The benchmark is not a hosted blind test set: ground truth is public in the code repository, so do not place groundtruth/ or verification/ files in a model's context.

Provenance and limits

The SOPs are sanitized executable Workato recipes; customer-identifying information was removed before benchmark artifacts were derived. Task descriptions and runtime variants were produced by model-based generation pipelines followed by five rounds of author review (task descriptions) and execution-based plus manual validation (variants). The benchmark targets structured enterprise workflows over 64 tool interfaces (Slack, Jira, Salesforce, Google Sheets, Google Drive, Snowflake, Workato and others) and should not be read as evidence that a model can safely run arbitrary workflows against live business systems.

Source files (source/)

Raw benchmark fixtures used by the simulator and scorer, packed as archives that unpack to their original paths in the code repository. Fetch and verify them with python scripts/download_data.py --only benchmark from a clone of the code repository (checksums in scripts/data_release.json).

archive	unpacks to	files
source/benchmark_input.tar.gz	benchmark/input/recipes/ (model-visible inputs per recipe)	7,120
source/benchmark_groundtruth.tar.gz	benchmark/groundtruth/recipes/ (reference per variant, scorer only)	3,220
source/benchmark_verification.tar.gz	benchmark/verification/{dataset,evaluation_support,groundtruth,seeds}/ (evaluator-private state)	12,262

Do not place groundtruth/ or verification/ contents in a model's context.

Citation

@article{agentorch2026,
  title   = {AgentOrch: How Close Can Global Orchestration Come to Step-by-Step Interaction?},
  author  = {Liu, Zijie and Huang, Qi Yang and Duan, Jinhao and Lee, Yueh Che and Joseph, Thomas A. and Turner, Zayne and Seligman, Adam and Chen, Tianlong and Xu, Zhaozhuo},
  year    = {2026}
}