AgentGA: Evolving Code Solutions in Agent-Seed Space

TL;DR

AgentGA is a novel framework that evolves autonomous code-generation by optimizing initial task prompts and archives, demonstrating superior performance in AutoML benchmarks and competitive tournament results.

Contribution

It introduces agent-seed optimization with a population-level genetic algorithm coupled with long-horizon agents, outperforming existing methods in AutoML tasks.

Findings

AgentGA achieves 71.90% Exceeds % of Human on Weco-Kaggle Lite benchmark.

Descendants with inherited parent archives win 51.9% of tournaments.

AgentGA outperforms the AIDE reference with 51.38% versus 15/16 competitions won.

Abstract

We present AgentGA, a framework that evolves autonomous code-generation runs by optimizing the agent seed: the task prompt plus optional parent archives that initialize a fresh workspace. The outer loop searches over these reusable starting conditions rather than editing code directly. Each generation launches a fresh autonomous run in an isolated workspace, while selected parent archives provide inherited artifacts that descendants can inspect and reuse. AgentGA couples a population-level genetic algorithm with long-horizon agents; selection uses deterministic 1:1 elite tournaments and operator allocation is adapted online with a modified Hedge controller. We instantiate the approach for tabular AutoML on the 16-competition Weco-Kaggle Lite benchmark. Across the full benchmark, AgentGA averages 71.90% Exceeds % of Human versus 51.38% for the AIDE reference, winning 15/16 competitions. Within AgentGA runs, descendants conditioned on inherited parent archives win 51.9% of 1,680 parent-child tournaments versus 8.6% for de novo proposals. These results support agent-seed optimization as a practical design choice for autonomous code-search systems.