Meta's AIRA₃ Agent Beats Human Coders in Kaggle Contest

Meta's Fundamental AI Research team recently entered its next-generation autonomous research agent, AIRA₃, into a live Nvidia-run Kaggle competition. Competing against roughly 4,000 human teams, the agent secured eighth place to win a gold medal. The challenge required participants to fine-tune a 30B Nemotron model to improve its reasoning capabilities, with submissions evaluated on a hidden, private test set.

Unlike traditional systems that rely on a central controller, AIRA₃ coordinates multiple long-running agents asynchronously. These agents communicate through a shared forum to exchange hypotheses and a shared filesystem to store solution artifacts. The live, gold-medal-winning run utilized an ensemble pairing GPT 5.5 with OpenCode alongside Claude 4.8 with ClaudeCode. In post-hoc evaluations on the same private test set, an ensemble of Muse Spark 1.2 with MuseCode also achieved gold, while Muse Spark 1.1 with OpenCode and GLM 5.2 with OpenCode both earned silver.

To test its adaptability, Meta applied the same architecture to other domains by changing only the task prompts. AIRA₃ achieved a 27% latency reduction on production GPU kernels and gold-level performance translating 4,000-year-old Akkadian clay tablets into English. Its predecessor, AIRA₂, is already a state-of-the-art baseline on MLE-bench-30, which spans 5 low, 20 medium, and 5 high complexity tasks. However, running these systems is resource-intensive; AIRA₂ required eight Nvidia H200 GPUs, and while the older aira-dojo framework is open source, Meta has not yet released AIRA₃.

For machine learning practitioners, this milestone shifts the focus of autonomous engineering from single-agent prompting to decentralized coordination. By allowing independent agents to build on each other's discoveries through a shared forum, the system bypasses the diminishing returns typical of long-horizon agent runs. It suggests that future breakthroughs in automated development will depend less on scaling monolithic models and more on designing effective collaborative environments for specialized coding agents.