Gaia2 and ARE: Empowering the community to study agents

Gaia2 is a new agentic benchmark that builds upon its predecessor, GAIA, to evaluate more complex AI agent behaviors. It tests agents on read-and-write tasks involving instruction following, ambiguity handling, and adaptability in a noisy environment with controlled failures. Released alongside Gaia2 is the open Meta Agents Research Environments (ARE) framework, designed to run, debug, and evaluate agents in these complex, real-world-like simulations. The benchmark includes new task groups such as time-sensitive reasoning and agent-to-agent collaboration to better reflect the challenges of open-world agent deployment.