A computational grid is a wide-area computing environment for cross-domain resource sharing and service integration. Resource management and load balancing are key concerns when implementing grid middleware and improving resource utilization. Grid resource management can be implemented as a multi-agent system with resource advertisement and discovery capabilities if job requests from users are associated with explicit QoS requirements. In this work agent-based self-organization is proposed to perform complementary load balancing for batch jobs with no explicit execution deadlines. In particular, an ant-like self-organizing mechanism is introduced and proved to be powerful to achieve overall grid load balancing through a collection of very simple local interactions. A modeling and simulation environment is developed to enable performance of the ant algorithm to be investigated quantitatively. Simulation results included in this work illustrate the impact of different performance optimization strategies on the overall system load balancing level, speed and efficiency.