ByteDance's Seed AI research team has identified a new scaling law showing that AI agents can double their learning speed every three months by interacting with real-world environments — a discovery that could extend the AI development runway at a moment when conventional training methods are running into hard limits. The findings were published in a research paper on Thursday, 3 July 2026.

What the research found

The Seed AI team at TikTok parent ByteDance demonstrated that autonomous software agents — systems that execute tasks on a human's behalf — improve at a measurable, predictable rate when deployed in real-world settings over extended periods. The team characterised this improvement curve as a new scaling law, distinct from the data- and compute-heavy approach that has defined AI progress for the past decade.

To benchmark this behaviour, the researchers developed EdgeBench, a suite of 134 ultra-long-horizon tasks spanning domains including software engineering, scientific discovery, formal mathematics, and professional knowledge work. Critically, each task requires a minimum of 12 hours of continuous AI agent operation — far beyond the short-burst evaluations common in existing benchmarks.

Why it matters

The global AI industry has been under growing pressure to find new paths to model improvement. Prominent figures including OpenAI co-founder Andrej Karpathy have publicly warned that the brute-force method of scaling up training data and compute cannot continue indefinitely. Separately, US-based research institute Epoch AI recently cautioned that publicly available, human-generated text data could be exhausted within the next six years.

Against that backdrop, the agentic AI paradigm — where models learn by doing rather than by ingesting static datasets — has attracted intense industry interest. However, the ByteDance researchers noted in the paper that how autonomous systems 'learn from real-world environments after deployment remains far less understood,' underscoring the significance of their findings.

The competitive backdrop

The discovery positions ByteDance alongside a cluster of labs racing to define the next phase of AI scaling. OpenAI, Anthropic (maker of Claude Opus), DeepSeek, and Zhipu AI are all investing heavily in agentic capabilities, each pursuing different approaches to post-training improvement. ByteDance's EdgeBench framework could become an industry-standard evaluation tool if adopted broadly, giving the company an outsized role in shaping how agent progress is measured.

What's next

The publication of EdgeBench as an open benchmarking suite invites independent replication and stress-testing by the wider research community. If the three-month doubling rate holds across diverse agent architectures and task categories, it would provide the AI industry with a credible, data-driven alternative to the stalling pre-training scaling curve. Investors and developers will be watching whether frontier labs adopt EdgeBench as a standard metric — and whether ByteDance's own models are the first to benefit from the methodology.