AutoDesign Framework: Optimizing Agent Harness, Beats Claude Design in Poster Generation
KyeGomezB · x · 2026-08-18
The paper introduces AutoDesign, a framework where a meta-harness optimizer guides a code agent to recursively improve the harness based on rollout feedback, rather than tuning model weights.
- Core Mechanism: Features an inner loop to fix artifacts and an outer meta-loop that learns from failures to update the harness itself.
- Performance: In 7 agent-model setups, the learned harness boosts the PosterBench score by an average of 12.4% (+5.0 to 19.6 points).
- Benchmark: AutoDesign achieves a score of 78.32 on the Main Track, surpassing the commercial system Claude Design by 7.45 points.
- Cost: The fully autonomous loop (40 mins, 253 tool calls) costs under $3, achieving average conference poster quality in human evaluation.
More from coding & agent
- Agent Governance Shifts to Device Level with mimOE Engine Release — shashib · 2026-08-18
- One prompt, ~2 hours: Claude + Thrixel autonomously build a full 3D game — RanaHanocka · 2026-08-18
- Permix: Lightweight Type-Safe TS Permissions Library Wins Developer Praise — jonathan_wilke · 2026-08-18
- Give agents a centralized source of truth before they hallucinate — eptwts · 2026-08-18
- After letting her agents queue up commits, her GitHub daily streak exploded to 2026 — christine_hall · 2026-08-18
- Why logging is vital for agent apps: accountability, evals, and self-healing — jasonkneen · 2026-08-18