In a research paper published Thursday, ByteDance’s Seed AI team revealed that AI agents — autonomous software that executes tasks on behalf of humans — can double their learning speed every three months by interacting with real-world environments over long periods.
This result comes as the global AI industry looks for new ways to improve models. For many years, developers have relied on feeding systems more data and computing power during initial training, but prominent industry figures – including Andrei Karpathy, co-founder of OpenAI – have warned that this brute force approach cannot continue forever.
However, despite the fact that tech companies are moving toward agentic AI, ByteDance researchers point out in the paper that how “these autonomous systems learn from real-world environments post-deployment is still largely poorly understood.”
To address this problem, the team developed EdgeBench, a benchmark suite of 134 very long tasks covering a wide range of domains from software engineering and scientific discovery to formal mathematics and professional knowledge work. Each task requires at least 12 hours of continuous operation of the AI agent.
