Meta’s AIRA₃: An Autonomous AI Research System That Lets Agents Learn from Each Other
Key Info
Meta’s autonomous AI research system AIRA₃, recently tested in an NVIDIA-run Kaggle competition, does not rely on a central controller. Instead, it coordinates many long-running agents through a shared discussion forum and a shared filesystem, allowing them to build on each other’s discoveries.
Highlights
- AIRA₃ runs many agent pairs (each a model plus coding harness) in separate isolated environments, coordinating asynchronously rather than through a single orchestrator.
- Two shared substrates enable collaboration: a forum for hypotheses and findings, and a filesystem for solution artifacts.
- This design lets compute compound knowledge: one agent’s discovery can be picked up and extended by others, driving overall performance.