MILO: Automated Harness Discovery via Orchestrated Multi-Agent Evolution

They introduce MILO (Meta-evolutionary Island Orchestration), a framework that co-evolves agent harnesses and the strategy used to discover them. MILO combines hierarchical lineage memory over island-based trees using rejected mutations as negative evidence, per-island mutator agents that rewrite complete harnesses using global history and parent feedback, and an orchestrator adapting search through lineage grafting and speciation. Across Terminal-Bench 2.1, PaperBench, and DeepSWE, MILO-discovered harnesses outperform eight state-of-the-art harnesses and six search methods. With Opus 4.8, MILO improves resolution over its initial harness by +12.0%, +28.3%, and +10.3%, respectively. On Terminal-Bench 2.1, it achieves 86.1 pm 2.0% while using 26% fewer tokens than its initial harness.

With Opus 4.8, MILO improves resolution over its initial harness by +12.0%, +28.3%, and +10.3%, respectively.

On Terminal-Bench 2.1, it achieves 86.1 pm 2.0%, exceeding the official leaderboard's top entry of 83.8 pm 2.3%.

Uses 26% fewer tokens than its initial harness on Terminal-Bench 2.1.