The research division has released data indicating that autonomous agents can accelerate their learning process by continuously performing real-world tasks within user and external system interaction environments.
In a newly published scientific study, ByteDance’s Seed AI team claims that AI agents—autonomous software systems designed to perform tasks on behalf of humans—demonstrate the ability to double their learning speed approximately every three months when engaged in extended interactions with real-world environments.
This does not refer to conventional training on pre-prepared datasets, but rather to a post-deployment process in which the agent improves its behavior through accumulated experience in actual usage scenarios. This approach is viewed as an alternative to the traditional strategy of scaling models by increasing data volume and computational resources.
The authors note that the AI industry is already encountering the limits of “brute force” in model training. Previously, industry representatives, including co-founder Andrej Karpathy, pointed out that simply scaling up computing power and data cannot remain the sole driver of progress in the long term.
At the same time, the researchers emphasize that the behavior of AI agents after deployment in real-world environments remains insufficiently studied, despite the growing shift of companies toward agent-based systems capable of handling complex, multi-step operations.
To analyze this process, the ByteDance team developed a benchmark called EdgeBench, which includes 134 lengthy tasks, each requiring at least 12 hours of continuous operation by the AI agent.
The tasks span a wide range of domains, including software engineering, scientific research, formal mathematics, and professional analytical work. This approach makes it possible to evaluate not only the accuracy of solutions but also the system’s ability to maintain stability and effectiveness during prolonged autonomous operation.
The findings align with a broader trend in the development of agent-based AI, where the key focus is not solely on the quality of the initial model, but also on its capacity to learn and adapt after being deployed in real-world operations.
About the Author
