Agentic ai //
Challenge & train frontier agents in realistic enterprise company simulations
Snorkel's collection of enterprise environments that simulate real-world companies, built by domain experts to train AI agents on enterprise workflows. Each environment includes thousands of tau-bench style tasks, complete with databases, company policy documents, domain tools, and simulated users that reflect the complexities agents have to operate in.
Snorkel has helped models improve from 10.9% to 42.0% pass@1 (Insurance Underwriting).
Developed by Snorkel’s AI Data Research Lab, each environment is authored and validated by teams of domain experts and calibrated against frontier models.
- Underwriting (Property & Casualty Insurance Company)
- Finance (Investment Banking / Equity Research)
Enterprise Environments are intentionally calibrated to stress-test state-of-the-art agents
Built for frontier model evaluation:
- Tiered difficulty across Base, Core, Advanced, and Frontier
- Frontier tasks where leading models score <40% accuracy
- Evaluated against GPT-5.4 and Claude Opus 4.6 in the OpenEnv harness
If your agent succeeds here, it performs in production.
Why the Snorkel Data Series
Expert-led validation
Every task is built and validated through a multi-layer quality pipeline.
Expert review
Subject-matter experts review every task for oracle correctness, prompt clarity, environment correctness, and tagging.
Programmatic checks
Automated validation ensures every task manifest is complete and not a trivial variation of an existing task.
Distribution guardrails
New submissions are accepted only if they maintain dataset-level balance across tool domains, difficulty levels, and categories.
Judge stability
LLM-as-judge scoring uses fixed parameters and is benchmarked against human annotators to control variance.
