AgentGA: Evolving Code Solutions in Agent-Seed Space
Researchers have introduced AgentGA, a novel framework designed to evolve autonomous code-generation processes by optimizing the 'agent seed,' which consists of task prompts and optional parent archives. Unlike traditional methods that directly edit code, AgentGA’s outer loop searches for reusable starting conditions. Each generation initiates a fresh autonomous run in an isolated workspace, allowing descendants to inspect and reuse artifacts from selected parent archives. The system integrates a population-level genetic algorithm with long-horizon agents, employing deterministic 1:1 elite tournaments for selection and a modified Hedge controller for online operator allocation. When tested on the 16-competition Weco-Kaggle Lite benchmark for tabular AutoML, AgentGA achieved an average of 71.90% Exceeds % of Human, significantly outperforming the AIDE reference model's 51.38%. It won 15 out of 16 competitions. Furthermore, descendants utilizing inherited parent archives won 51.9% of parent-child tournaments, compared to only 8.6% for de novo proposals. These findings highlight agent-seed optimization as a practical and effective design strategy for enhancing autonomous code-search systems.
Wire timeline
AgentGA: Evolving Code Solutions in Agent-Seed Space
Researchers have introduced AgentGA, a novel framework designed to evolve autonomous code-generation processes by optimizing the 'agent seed,' which consists of task prompts and optional parent archives. Unlike traditional methods that directly edit code, AgentGA’s outer loop searches for reusable starting conditions. Each generation initiates a fresh autonomous run in an isolated workspace, allowing descendants to inspect and reuse artifacts from selected parent archives. The system integrates a population-level genetic algorithm with long-horizon agents, employing deterministic 1:1 elite tournaments for selection and a modified Hedge controller for online operator allocation. When tested on the 16-competition Weco-Kaggle Lite benchmark for tabular AutoML, AgentGA achieved an average of 71.90% Exceeds % of Human, significantly outperforming the AIDE reference model's 51.38%. It won 15 out of 16 competitions. Furthermore, descendants utilizing inherited parent archives won 51.9% of parent-child tournaments, compared to only 8.6% for de novo proposals. These findings highlight agent-seed optimization as a practical and effective design strategy for enhancing autonomous code-search systems.
cs.AI updates on arXiv.org