Amazon's MILO auto-evolves agent harnesses, hitting 86.1% on Terminal-Bench 2.1
amazon · hf · 2026-10-02
Amazon introduces MILO (Meta-evolutionary Island Orchestration), a framework that automatically discovers agent harnesses — the outer systems controlling model execution and environment interaction, which strongly affect long-horizon performance but demand heavy manual tuning.
How it works:
- Hierarchical lineage memory over island-based trees, using rejected mutations as negative evidence;
- Per-island mutator agents that rewrite complete harnesses using global search history and parent-specific feedback;
- An orchestrator adapting search via lineage grafting, speciation, mutator reassignment and curriculum revision.
Results: Across Terminal-Bench 2.1, PaperBench and DeepSWE, MILO-discovered harnesses beat eight SOTA harnesses and six search methods. With Opus 4.8 it improves resolution over the initial harness by +12.0%, +28.3% and +10.3%; Terminal-Bench 2.1 reaches 86.1±2.0%, topping the official leaderboard (83.8±2.3%) while using 26% fewer tokens. It also advances best-known bounds on EinsteinArena open math problems.
More from coding & agent
- Generic API-catalog MCP server vs one tool per job: a context-bloat experiment — MandeepKaur9407 · 2026-11-16
- Berkeley paper: LLMs answer with stale values even when the update is in context — ombre-ai · 2026-11-16
- skills-gate: one gate skill fronts 148 agent skills at just 389 tokens — Distinct-Pie2389 · 2026-11-16
- Open-weight Karpo-Decision-31B tops Decision Index at 73.9, one forward pass per answer — CodeByPoonam · 2026-11-16
- shadcn boosts @uisdk: one-shot UI rebuild that's instantly E2E-testable by AI — shadcn · 2026-11-16
- AI agent deletes 43,599 lines of OpenClaw test code in one PR — vincent_koc · 2026-11-16