MILO: Automated Harness Discovery via Orchestrated Multi-Agent Evolution
MILO searches for whole agent harnesses, and also changes the search strategy while it runs.
The paper says harness design is a major driver of long-horizon agent performance, but hard to tune manually as models change. MILO uses island-based lineage memory, mutator agents, and an orchestrator that can graft lineages, split species, reassign mutators, and revise curricula. In reported tests on Terminal-Bench 2.1, PaperBench, and DeepSWE, its discovered harnesses beat the compared harnesses and search methods across frontier and open-weight models. On Terminal-Bench 2.1 with Opus 4.8, it reports 86.1 ± 2.0% while using 26% fewer tokens than its initial harness. HF Daily Papers' note
The paper says harness design is a major driver of long-horizon agent performance, but hard to tune manually as models change. MILO uses island-based lineage memory, mutator agents, and an orchestrator that can graft lineages, split species, reassign mutators, and revise curricula. In reported tests on Terminal-Bench 2.1, PaperBench, and DeepSWE, its discovered harnesses beat the compared harnesses and search methods across frontier and open-weight models. On Terminal-Bench 2.1 with Opus 4.8, it reports 86.1 ± 2.0% while using 26% fewer tokens than its initial harness. HF Daily Papers' note
score 4