UndoBench

UndoBench is a benchmark for evaluating whether tool-using AI agents can recover safely from execution failures without duplicating, losing, or corrupting externally visible effects. The benchmark separates task competence from recovery capability using paired CONTROL and FAULT executions. It includes 36 enterprise workflows across 8 domains, with explicit fault scenarios and state/effect oracles for evaluation.

Benchmark at a Glance

Workflows	36
Domains	8
DEV	14
Validation	10
Test	12
Benchmark version	v1.0.1

Quickstart

from datasets import load_dataset

ds = load_dataset("TanmaySah/undobench")

print(ds)
print(ds["test"][0])

What Does One Row Represent?

Each row represents one canonical UndoBench workflow specification.

The 36 rows are benchmark task definitions, not individual model executions or aggregate experimental results.

Each workflow includes the task objective, available tools, fault metadata, environment/evaluation metadata, and provenance.

Full benchmark execution code and reproducibility artifacts are available in the GitHub repository.

Schema

Column	Type	Description
task_id	string	Unique workflow identifier
split	string	dev / validation / test
domain	string	Enterprise domain
workflow_name	string	Canonical workflow title
workflow_description	string	Baseline workflow description
objective	string	Agent task instruction
fault_scenario	string (JSON)	Fault-injection specification
fault_boundary	string	Failure lifecycle boundary
reversibility	string	Action reversibility class
idempotency_support	bool	Native idempotency support
tools	list[string]	Available tools
initial_state	string (JSON)	Evaluator environment setup
expected_outcome	string (JSON)	Expected final-state assertions
required_effects	string (JSON)	Required semantic effects
oracle_spec	string (JSON)	Evaluation oracle
benchmark_version	string	Benchmark version (v1.0.1)
source_path	string	Canonical implementation path

Evaluation note: initial_state, expected_outcome, required_effects, oracle_spec, and fault-injection metadata are published for reproducibility and are not provided to evaluated agents as privileged runtime inputs.

Evaluation

The executable UndoBench evaluation harness is available in the GitHub repository.

Install and verify the benchmark:

git clone https://github.com/tradertanmay/undobench.git
cd undobench
pip install -e ".[all]"
undobench doctor
undobench smoke

Evaluate a custom agent on the validation split:

undobench run \
  --agent my_agent.py:MyAgent \
  --split validation \
  --recovery none \
  --seeds 42 \
  --output runs

Or evaluate an OpenAI-compatible model endpoint:

undobench run \
  --model openai/<model-name> \
  --split validation \
  --recovery naive \
  --seeds 42 \
  --output runs

After a run, recompute metrics and generate a standardized report:

undobench evaluate runs/<run_id>
undobench report runs/<run_id>

UndoBench reports task competence and recovery metrics including Control Pass Rate, RSR, CRSR, EOR, DER, and MER.

The frozen test split requires explicit opt-in before execution. See the GitHub repository for full evaluation and integration instructions.

Paper & Code

• Paper (arXiv)
• Hugging Face Paper
• GitHub Repository

Citation

@article{undobench2026,
  title={UndoBench: Separating Task Competence from Recovery Capability in Tool-Using AI Agents},
  author={Dolly Sah and Tanmay Sah and Harshul Jain and Tanya Sah},
  journal={arXiv preprint arXiv:2610.05622},
  year={2026},
  url={https://arxiv.org/abs/2610.05622}
}

License

Dataset and benchmark specifications: CC BY 4.0

Software and execution framework: Apache 2.0