Research platform
ARE
Meta Agents Research Environments
Build and simulate environments with applications, tools, events, and verifiers. ARE supports research into agents that must reason through multi-step tasks and adapt as the world changes.
Open research · Environments & benchmarks
Build environments.Evaluate agents.
Explore the tools and benchmarks for studying how AI agents reason, act, and adapt in dynamic environments.
Explore the projects
Agent benchmark
GAIA2
Evaluate agents in a changing world
A benchmark built in ARE for general-purpose assistants. Its scenarios test execution, search, adaptability, temporal reasoning, ambiguity, collaboration, and resilience to noise in simulated environments.
Container-based evaluation
GAIA2-CLI
Evaluate agent runtimes end to end
An evaluation stack that packages GAIA2 application CLIs and runs agents in containers. It grades tool use against scenario reference behavior and produces inspectable traces of each run.