Tsinghua's Design Anything uses intent-execution graphs to keep design agents on-brief at scale
Xuchao Ren · hf · 2026-10-22
A Tsinghua/MIT Senseable City Lab/Alibaba DAMO paper introduces Design Anything, a graph-based harness for design agents: an intent-execution graph preserves design intent and binds deliverables to workflow steps, while task contracts retain constraints and a deterministic controller checks outputs and tracks outstanding deliverables.
DesignAnything-Bench: 228 tasks across 19 design domains at 6/14/24-artifact scales. With the LLM fixed, it satisfies 85.0% (DeepSeek V4.1 Flash High) and 92.5% (GPT-5.6 Sol High) of factual constraints, beating strongest same-model baselines by 15.4% and 14.4%, and by 25.1%/23.0% on 24-artifact tasks.
More from coding & agent
- One Reply to Grok Bot Auto-Built a Full Starship Flight Tracker Dashboard — XFreeze · 2026-10-22
- Researcher details his 100k-line personal agent system built around Obsidian and Telegram — s_scardapane · 2026-10-22
- Teknium's Hermes Labs plugins let Hermes learn browser workflows by watching you — Teknium · 2026-10-22
- In test, Copilot found MetricDescriptor first try; Cursor and o1 needed repeated corrections — HaktanSuren · 2026-10-22
- Using the moov agent to generate and trim video clips straight from the terminal — wavefnx · 2026-10-22
- YC-backed Kiwi AI launches as a data engineer for GTM agents — ycombinator · 2026-10-22