What is Gaia2?

Gaia2 is Meta’s test of a personal-assistant agent operating a simulated phone, contacts, messages, and calendar included, where the clock keeps running so a message can arrive or a meeting can move while the model is still thinking. It replaced the original GAIA because that benchmark’s web-browsing tasks had gotten saturated and never tested an agent adapting to a changing situation in the first place.

Gaia2 runs on Meta’s Agents Research Environments (ARE), an open-source platform released on GitHub under the MIT license, so a team isn’t limited to Gaia2’s own 1,120 scenarios; ARE lets you build a custom dynamic environment and run the same time-pressure, ambiguity, and tool-failure conditions against your own agent. A verifier checks every action the agent took against what should have happened, agreeing with human graders 98% of the time. Does passing an eval suite mean an agent stays reliable covers why a strong Gaia2 score still isn’t a production guarantee.