AutoDesign: Self-Evolving Harness for Long-Horizon Agentic Design
机器之心 · wechat · 2026-08-21
AutoDesign introduces a Meta-Harness Optimization framework to iteratively improve the Harness (process orchestrator) in multi-agent systems. Using a dual-loop architecture, the inner loop executes design tasks while the outer loop analyzes trajectories to identify failure patterns and update Harness components via a Code Agent (modifying one component at a time to ensure credit assignment). Validated on a "Paper-to-Poster" task with the proposed PosterBench (100 papers), AutoDesign's score improved from 49 to 80.9 through self-evolution. It outperforms Claude Design by 7.45 points under equivalent settings and boosts Coding Agent performance by an average of 12.4 points.
More from coding & agent
- PuppyOne Editor: Windows XP-Style Desktop Integrated with AI Agents — Aiden_Tech_Ai · 2026-08-21
- Code quality is dead; it's all about tooling quality now — madhavsinghal_ · 2026-08-21
- Semantic Caching Test: Rewriting Bad Cache Hits Is Worse Than Missing — Reasonable_Royal_621 · 2026-08-21
- OpenCode Offers Free Access to Ox Alpha: 1M Context and Multimodal — TJLarkin23 · 2026-08-21
- Best practices for AI agents: narrow, atomic, explainable — with a 4-agent hedge fund demo — arieljalali · 2026-08-21
- Agent Routing Progress: Intent Classification and Skill Routing Work, Data Permissions Lag — arieljalali · 2026-08-21