---
title: "AutoDesign: Meta-Harness Optimization for Long-Horizon Agentic Design"
source: "arXiv AI/CL/LG（论文）"
source_tier: "t2"
url: "https://arxiv.org/abs/2608.13560v1"
permalink: "https://feed.astralor.com/items/itm_b8c1eb736a42ef3959de6e87"
published: "2026-08-13T17:59:57.000Z"
score: 76
channel: "ai"
tags: ["AutoDesign","Claude Design","DesignHarness","PosterBench","Meta-learning","Agent"]
---

# AutoDesign: Meta-Harness Optimization for Long-Horizon Agentic Design

## AI commentary

> This paper stands out for turning 'agent self-improvement' from a slogan into a quantifiable, verifiable framework, and generating high-quality posters for under $3 is especially attractive for resource-constrained teams. For those building agent products, automated design tools, or content pipelines, this methodology is likely directly adaptable.

## At a glance

AutoDesign presents a meta-optimization framework that leverages a code agent to recursively improve the model-harness system based on rollout feedback, enabling self-improving long-horizon agentic design. The paper focuses on academic paper-to-poster generation and introduces PosterBench, a benchmark comprising 100 papers across five disciplines. AutoDesign achieves a score of 78.32 on PosterBench Main Track, surpassing the closed-source commercial system Claude Design by 7.45 points. Across seven controlled code-agent-model configurations, integrating the learned DesignHarness consistently improves performance, increasing the average PosterBench Score from 54.99 to 67.39 (+12.4%). In a fully autonomous long-horizon loop, it executes 253 tool calls and 11 editing turns within 40 minutes for under $3, reaching average conference-poster quality in human evaluation. A system-blind human study further demonstrates that AutoDesign achieves the highest human preference among evaluated systems.

## Key points

- Proposes AutoDesign, a meta-optimization framework where a code agent recursively improves the model-harness system based on rollout feedback for long-horizon self-improvement.
- Introduces PosterBench, a 100-paper Main Track spanning five disciplines plus a shared 10-paper mini subset for controlled evaluation.
- AutoDesign scores 78.32 on PosterBench Main Track, surpassing Claude Design by 7.45 points; the learned DesignHarness improves average score from 54.99 to 67.39 (+12.4%) across seven configurations.
- In a fully autonomous loop, it executes 253 tool calls and 11 editing turns within 40 minutes under $3, reaching average conference-poster quality in human evaluation.

## Full text

Transforming multimodal sources into condensed and structured media outputs can be fundamentally conceptualized as a long-horizon agentic process centered on a model-harness system. While an ideal harness system should align with human design priors and accumulate reusable experience through empirical exploration to drive recursive self-improvement, existing paradigms remain static and fall short of this capability. In this paper, we present AutoDesign, a framework that aligns with human design priors, where a meta-harness optimizer guides a code agent to recursively improve harness based on rollout feedback. To instantiate and evaluate this framework, we focus on the academic paper-to-poster generation task and introduce PosterBench, comprising a 100-paper Main Track spanning five disciplines and PosterBench-mini, a shared 10-paper subset for controlled evaluation. On the PosterBench Main Track, AutoDesign achieves the highest score of 78.32, surpassing the closed-source commercial system Claude Design by 7.45 points. Across seven controlled code-agent-model configurations, integrating the learned DesignHarness consistently improves performance, increasing the average PosterBench Score from 54.99 to 67.39 (+12.4%). In a fully autonomous long-horizon loop, it executes 253 tool calls and 11 editing turns within 40 minutes for under $3, reaching average conference-poster quality in human evaluation. A system-blind human study further demonstrates that AutoDesign achieves the highest human preference among evaluated systems.

---

Verify against the original before citing; curated by Astral Feed: https://feed.astralor.com/items/itm_b8c1eb736a42ef3959de6e87