Meta AI introduces MIRA architecture for long-horizon research agents
Meta AI researchers introduced MIRA, a dual-component agent architecture that decouples meta-reasoning from execution to improve decision-making in long-horizon research tasks.
Meta AI researchers introduced MIRA, a dual-component agent architecture that decouples meta-reasoning from execution to improve decision-making in long-horizon research tasks.
DeepLearning.AI reports that Meta AI uses dedicated memory agents to record tools and past errors and insert short reminders, raising Claude Sonnet 4.5 benchmarks from 37.6% to 45.9%.