A new development in artificial intelligence has emerged from a collaboration between researchers at UNC-Chapel Hill, Salesforce Research, and Stanford University. They have introduced a groundbreaking framework called Agent0. This system is designed to create high-performing AI agents without relying on large human datasets. Instead, it teaches itself through a process known as multi-step co-evolution, incorporating tools seamlessly into its learning.
Agent0 focuses on improving mathematical and general reasoning skills. The researchers found that by carefully generating tasks and integrating tool usage, they could push the capabilities of a base model beyond its original limits. They tested Agent0 across ten benchmarks and reported promising results.
The framework operates using two distinct agents. The first is a Curriculum Agent, which generates tasks for learning, while the second is an Executor Agent that solves these tasks using a Python tool. The training process involves two main stages: the Curriculum Agent evolves by creating tasks, while the Executor Agent learns to solve them. This back-and-forth process allows both agents to improve over time.
To assess how well the Curriculum Agent creates tasks, the researchers use a composite reward system. This system takes into account three factors: how uncertain the tasks are, how often the executor uses the tool, and a penalty for generating repetitive tasks. This encourages the Curriculum Agent to design challenges that are difficult yet solvable.
The Executor Agent learns from its own outputs rather than relying on pre-existing answers. It generates a pool of candidate tasks, filtering them based on how well it can handle them. This creates a set of challenging tasks that avoid being too easy or too difficult. The training method, called Ambiguity Dynamic Policy Optimization (ADPO), helps stabilize the learning process by adjusting how the agent responds to uncertain tasks.
The results of Agent0 have been impressive. It was tested on various benchmarks for mathematical and general reasoning, including well-known assessments like AMC and MATH. On the Qwen3 8B Base model, Agent0 achieved an average score of 58.2 in math, significantly higher than the 49.2 score of the base model. It also improved general reasoning scores from 34.5 to 42.1.
Agent0 has outperformed other data-free AI frameworks, such as R Zero and Absolute Zero, demonstrating its effectiveness in evolving without external data. The researchers noted that the tasks generated by the Curriculum Agent evolved from simple questions to more complex problems, showing a clear progression in difficulty.
This development marks a significant step toward creating AI that can learn and improve independently. Agent0 illustrates how combining co-evolution and tool integration can result in better-performing AI agents. As this technology advances, it could lead to more capable and versatile AI systems in the future.