BenchFlow is a frontier environment lab for AI agents. It builds stateful environments with services, files, tools, verifiers, traces, and replay so agents can learn from and be evaluated on whole workflows.
Products & media
A benchmark evaluating whether procedural skills help AI agents perform better across real work tasks.
An arena for contributing environments, post-training models, and scoring which capabilities generalize.
A local runtime for running real tasks with Docker, ACP agents, and independent verifiers.
A set of wire-compatible simulated workplaces that score agent capability and safety.
A robotics environment where coding agents control simulated robots across task suites.
An open benchmark evaluating whether AI agents can carry out end-to-end physics research.
Team
Founders
1 person
