Coding agents still need hand cleanup on most tasks, Co-Design Bench finds

Researchers from Alibaba and several Asian universities built Co-Design Bench, in which coding agents rework existing web apps based on user requests, sometimes using images, audio, video or documents as input. The tasks come from real coding sessions, with a preference for requests users had to rephrase or roll back, and range from small tweaks to full project overhauls. An AI judge tests the finished app in a browser, and a second judge rechecks every positive verdict for mock-ups or hardcoded results.

GPT-6 Astra leads the ten models tested, followed by Claude Opus 5 and Qwen3.8-Max, with Grok 4.6 in last place. Performance drops for every model when a task combines three or four types of change, such as layout, design, interaction and content. GLM-5.2 uses nearly six times as many tokens as GPT-6 Astra on smaller changes and still scores significantly worse.

The working environment also shifts results. Qwen3.8-27B scored either 75 or 56 percent on larger overhauls depending on the harness, and which of two models came out ahead changed with the environment. The authors recommend reporting the model and its harness together, and note that the test covers one fixed request per task, without follow-up questions or mid-task corrections.

If you build coding agents, plan for review and manual fixes on many real change requests rather than trusting a model's claim that the work is done. A leaderboard that lists only a model can mislead, so test the model and its harness as a pair on your own tasks. Requests that mix several kinds of change are where these agents fail most often.