The paper introduces *agent symbolic learning*, a framework that enables language agents to optimize themselves using symbolic optimizers. This framework is designed to transition from model-centric or engineering-centric approaches to data-centric learning, which is crucial for achieving artificial general intelligence (AGI). The key idea is to treat agents as symbolic networks where learnable weights are defined by prompts, tools, and their stacking. The framework mimics back-propagation and gradient descent from connectionist learning, but operates with natural language representations of weights, loss, and gradients. Experiments on standard benchmarks and complex real-world tasks demonstrate the effectiveness of agent symbolic learning, showing that language agents can update themselves after deployment, resulting in "self-evolving agents". The framework is open-sourced to facilitate future research on data-centric agent learning.The paper introduces *agent symbolic learning*, a framework that enables language agents to optimize themselves using symbolic optimizers. This framework is designed to transition from model-centric or engineering-centric approaches to data-centric learning, which is crucial for achieving artificial general intelligence (AGI). The key idea is to treat agents as symbolic networks where learnable weights are defined by prompts, tools, and their stacking. The framework mimics back-propagation and gradient descent from connectionist learning, but operates with natural language representations of weights, loss, and gradients. Experiments on standard benchmarks and complex real-world tasks demonstrate the effectiveness of agent symbolic learning, showing that language agents can update themselves after deployment, resulting in "self-evolving agents". The framework is open-sourced to facilitate future research on data-centric agent learning.