Researchers at ByteDance, the parent company of TikTok, have identified a new scaling law that describes how artificial intelligence (AI) agents can accelerate their learning by engaging in real-world tasks over extended periods. This discovery, detailed in a research paper published Thursday, could provide a fresh approach to advancing AI capabilities as traditional methods face increasing limitations.
The ByteDance Seed AI team found that AI agents—autonomous software systems designed to perform tasks on behalf of humans—can effectively double their learning speed every three months by continuous interaction with real-world environments. This contrasts with earlier development approaches that primarily focused on expanding training datasets and computational resources.
Experts have increasingly warned that relying solely on more data and computing power is unsustainable. OpenAI co-founder Andrej Karpathy has cautioned about the eventual plateau of improvements from these means. Moreover, a recent warning from the US-based research institute Epoch AI highlighted concerns about an impending shortage of publicly available, human-generated text data within six years, underscoring the urgency for alternative pathways to AI advancement.
Despite growing industrial interest in "agentic AI"—systems capable of autonomous decision-making and action—there remains limited understanding of how these agents learn and improve post-deployment in dynamic environments. To investigate this, the ByteDance researchers developed EdgeBench, a comprehensive benchmarking suite comprising 134 tasks that require AI agents to operate continuously for at least 12 hours. These tasks span diverse fields, including software engineering, scientific discovery, formal mathematics, and professional knowledge work.
In conducting over 38,000 hours of environment interaction, the team evaluated five cutting-edge AI models, including Anthropic’s Claude Opus 4.8, OpenAI’s GPT-5.5 and GPT-5.4, and Chinese AI systems Zhipu AI and DeepSeek. Their analysis revealed that AI agent performance during real-world interaction follows a predictable mathematical curve, indicating that capability improvements can continue steadily through hands-on experience even as conventional pre-training benefits diminish.
The researchers argue that "post-deployment learning from rich environments" should receive systematic attention comparable to the focus historically placed on pre-training processes. This ongoing adaptability is increasingly critical as AI agents are deployed across an expanding array of applications, from enterprise software to research and engineering tasks. Rather than relying on static knowledge acquired during training, these systems must evolve continuously to maintain effectiveness.
“An agent’s ability to learn from its environment and improve task performance is central to deploying AI systems at scale in the real world,” the ByteDance team concluded on the EdgeBench benchmark’s official platform. This insight could help sustain the momentum of AI development amid resource constraints and emerging technical challenges.